# Sahra-Josephine Hjorth — full essay corpus

> Complete plain text of every essay published on https://sahra-josephine.com, in English and Danish.

Attribution: quote as "Sahra-Josephine Hjorth" and link the English URL of the essay quoted.
Danish text is an automatic translation of the English original; cite the English page as the source.
Canonical entity: https://sahra-josephine.com/#person
Fact sheet: https://sahra-josephine.com/facts

## English (original)

### https://sahra-josephine.com/blog/everything-app

← Blog

AI & Society·September 8, 2026Elon Musk Bought the Wrong Real Estate for the Everything App

He spent years announcing that X would become the app for everything. OpenAI may have quietly built something closer to it first, because the most valuable real estate is not necessarily where we spend attention. It is where we express intent.

Silicon Valley has spent the last twenty years trying to put you inside The Truman Show. Not in the sense that Mark Zuckerberg wants to secretly cast your neighbors and film you while you sleep. In the more mundane, and probably more lucrative, sense that the largest technology companies have repeatedly tried to construct a world in which almost everything you do happens on property they own.

You talk there. You shop there. You watch television there. You book a holiday there. You order dinner there. You send money there. Ideally, you never have a reason to leave. That is the super-app dream.

Elon Musk has been unusually explicit about it. He bought Twitter for $44 billion, renamed it X and set out to add payments, video, creators, financial services and AI. He has described the ambition plainly: the everything app.

In late July, X Money launched for select users in the United States. You can send money to other users, get an X-branded Visa card, earn interest on deposits and start doing the kind of things that have absolutely nothing to do with arguing with strangers on Twitter. It is the latest very literal step in Musk’s years-long attempt to construct his own version of Seahaven.

I think he may have been right about the everything app. I just think he bought the wrong real estate. Because while Musk has been publicly announcing that he intends to build one, OpenAI has been doing something much less theatrical. It started with a chat box. Then the chat box learned to research. It learned to work with files. It got apps. It learned to shop. It can now navigate websites and take actions on a user’s behalf.

There was no grand announcement that ChatGPT intended to become the place where you do everything. It just kept absorbing things we previously opened other products to do. At some point, that stops looking like a chatbot with feature creep. It starts looking suspiciously like the everything app. Or at least a much stronger version of one. Maybe the company that wins the race to build the everything app does not need to own Truman’s entire world. It only needs to become the person he asks for everything.

Musk bought attention

The logic behind buying Twitter made sense. If you want to build an everything app, start with a place where hundreds of millions of people already spend time. Musk bought the property and began adding more things to it. Payments. Video. Creators. AI. Financial services. The strategy is essentially to keep building until there are fewer reasons to leave.

That is how we have historically thought about valuable internet real estate. Amazon owns the place I go when I want to buy something. LinkedIn owns the place I go when I want something related to work. Booking owns a destination for travel. YouTube owns video. Twitter owned the public conversation. The user decides where to go and then tells the product what they want. But AI reverses that sequence. Increasingly, I do not need to know where I should go first. I can start with what I am trying to accomplish.

I need somewhere to stay in Barcelona, but I am leaving for the airport at three. I need to understand whether this contract is terrible. My kid does not understand fractions. I want a pair of turquoise Aquazzura heels but not for $900. I need to know whether this company is actually growing or merely producing enthusiastic press releases. I can begin with the problem. The destination comes later.

That sounds like a small change in interface design. I think it is an enormous change in where power sits on the internet. X wants me to do everything there. ChatGPT only needs me to start everything there.

The best address on the internet may be intent

For most of the internet’s history, companies have spent extraordinary amounts of money trying to determine what we want. Advertising is largely an elaborate exercise in inference. I watched this. I clicked that. I searched for Mexico. I lingered on a sofa. Somewhere, several computers conclude that I may be moving house and spend the next month showing me dining tables. Google got much closer to the valuable part because search captures explicit intent. I type “hotel Mexico City,” and Google does not have to infer very much.

That distinction produced dramatically different economics. Twitter had enormous cultural relevance, yet generated approximately $4.5 billion in advertising revenue in 2021 and said it represented less than three percent of the digital advertising market. Google Search and related advertising generated $63.3 billion in the second quarter of 2026 alone. That is not a clean comparison. The companies differ in scale, geography, products and advertising infrastructure. But it illustrates why the moment at which somebody expresses a need has been such valuable territory.

Traditional search still largely returned the decision to me. Here are the links. Here are the hotels. Here are the advertisements. Good luck. The agentic version goes further. I tell the system the outcome I want, and it can increasingly research the options, compare them, recommend what to do and execute the next step.

Google monetized: What do you want to find? An agent can potentially monetize: What do you want done? That is the piece of real estate I think Musk underestimated: the moment before the user chooses the company.

I bring the engagement with me

But the economic comparison does not capture the strangest difference between X and an LLM. Traditional consumer platforms need somebody else to keep producing something worth engaging with. Facebook needs friends. X needs people saying things. YouTube needs creators. TikTok needs a relentless supply of videos capable of holding our attention. Without people, posts and videos, the property is empty. An LLM can produce a long engagement loop from something I bring with me.

I can arrive with a business problem, an unfinished idea, a decision I cannot make, a relationship I do not understand or something I am ashamed to ask another person. There does not have to be a new video waiting for me. Nobody else has to be online. Another person does not have to create the object of my attention before the session can begin. I bring the raw material for the engagement with me.

This is already visible in how people use the product. In OpenAI’s analysis of consumer ChatGPT conversations, 49 percent of messages were classified as “Asking”—seeking information, guidance or advice—while another 40 percent involved “Doing.” People were not primarily arriving to consume something another person had made. They were bringing questions, decisions and unfinished work from their own lives.

One model depends primarily on engagement with other people and the things they produce. The other can begin with an engagement between me, my own life and the model. That changes what the platform can know. Attention tells a platform what held me. Intent tells it what I want. Conversation can tell it why.

Meta has historically had to observe that I watched six videos about Milan and infer that I may want to go there. In a conversation, I might simply explain that I have ended a longstanding friendship, want to be somewhere lively, hate large corporate hotels, have overspent recently and do not want to repeat the last trip I took there. That is not behavioral exhaust from the engagement. It is the engagement.

The everything app may not contain everything

The traditional super-app thesis is about accumulation. Put messaging, payments, shopping, entertainment and services into one product until it contains such a large part of somebody’s life that leaving becomes inconvenient.

The AI version requires much less ownership. It can leave the airline as an airline and the hotel as a hotel. Shopify can still power the store. The bank can still hold the money. The learning company can still provide the credential. The underlying services do not have to disappear. They just become things the agent can choose.

This is no longer entirely theoretical. Product data from millions of merchants is already integrated into ChatGPT through Shopify Catalog. The merchant remains responsible for the store, brand, checkout and fulfillment. ChatGPT can occupy the conversation in which the customer describes the need and evaluates the options.

The everything app may not be where I do everything. It may be where I explain what I want done.

Consumer behavior is beginning to follow. Adobe found that traffic from generative-AI services to American retail websites increased 393 percent year on year in the first quarter of 2026. In March, those visitors converted 42 percent better than other traffic. The retailer still completed the transaction. But an increasing part of the decision had already happened somewhere else. The store remains. The visit becomes optional.

That is why the everything-app idea looks different through an AI lens. Musk’s version is about bringing more of the internet inside X. The agentic version can leave the internet where it is and become the place from which I access it. The browser asked where I wanted to go. Search asked what I was looking for. The agent asks what I am trying to accomplish. The underlying internet can remain enormous while becoming increasingly invisible to the person using it.

Attention is not obsolete

There is an obvious problem with declaring intent the superior real estate. Attention can create intent. I do not open Instagram, YouTube or X because I have already decided that I need a particular pair of shoes, restaurant or holiday. Sometimes I want the thing because somebody placed it in front of me.

Meta is the strongest possible argument against pretending attention is economically weak. It generated $196.2 billion in advertising revenue in 2025. But Meta built that machine by becoming exceptionally good at extracting inferred commercial intent from attention.

The closer it gets to knowing what I might buy, the more valuable the attention becomes. An LLM can skip much of that inference. The engagement itself contains the intent.

The distinction is therefore not that attention mattered yesterday and intent will matter tomorrow. Attention remains extraordinarily valuable for creating demand. Intent is extraordinarily valuable for allocating it. An agent that can act on the request potentially moves one step further again, from explicit intent to delegated intent. Feeds can help form desire. Agents can decide where it goes.

There is also an obvious complication: Musk owns an AI company too. Grok is deeply integrated into X, and there is no technological law preventing X from becoming much more useful when a user arrives with explicit intent. This is not an obituary for X. It is a question about habit. Can a destination built primarily around attention become the place people instinctively begin when they want something done? Or does the place they already ask to do things have an easier path into more of what they do?

And then there is money

X Money may be the strongest argument for Musk’s version of the everything app. Payments are unusually powerful infrastructure. Money touches almost everything. If X can become a place where people hold money, pay one another and use a card, it creates both extraordinary transaction frequency and a relationship that is considerably harder to leave.

That is not a trivial advantage. But it also exposes the difference between the two models. Musk’s version says: own the payment system too. The agentic version does not necessarily have to. If I ask an agent to buy something, the merchant can remain the merchant, the bank can remain the bank and the payment provider can remain the payment provider. The agent only needs permission to connect them. Which creates a more interesting question than whether payments matter. Is the more valuable position the one that owns the wallet? Or the one that receives the request and decides where the wallet should send the money? They are not the same bet.

The agent receives Milan and me

There is another reason being asked first matters. Intent is what I tell a system now. Context is what it already knows.

Imagine I ask it to find me somewhere to stay in Milan. A hotel website may know my dates, loyalty status, previous bookings and whatever preferences I have explicitly entered.

An assistant I have used for years can potentially know considerably more. It may know that I follow Formula 1 and love ballet. That I spend an unreasonable amount of time thinking about restaurants. That I care about fashion, generally stay in five-star hotels, belong to Soho House and have previously rejected recommendations because I thought the properties looked soulless. None of those facts is particularly valuable on its own. Together, they change the meaning of the request.

A booking site receives Milan and my dates. The agent receives Milan and me. But that creates the hardest objection to the entire argument. The interface that understands my intentions can also influence them. A search engine usually meets the request after I have reduced it to a query. An assistant may be present while I am still deciding what the request is.

If the same system helps me articulate a desire, determines which options deserve consideration, recommends one of them and can execute the transaction, advice and commercial influence become difficult to separate. The assistant may claim to understand what I want while its business model rewards it for changing where that desire goes—or even helping create it.

Owning contextualized intent is therefore not only more commercially valuable than owning attention. It may also be more dangerous. The old platforms watched us to infer who we were. The new interface invites us to explain ourselves directly.

Musk bought a property where other people must continually create reasons for me to visit. OpenAI may be building one where I bring the reason with me.

Musk is trying to build Seahaven: the place where everything happens. OpenAI may be building the place where everything begins.

Related essays

AI & Society · August 16, 2026AI Ate the Internet. Now We Want It to Decide What’s Human.Watermarking is supposed to make synthetic content transparent. It may also turn AI companies into the institutions we ask to certify human authorship.

AI & Edtech · September 5, 2026When Your Customer Becomes Your Competitor: Your Customer Found a Shortcut to the Software FactoryPart four of how 90% of edtech disappears. Software companies spent twenty years convincing customers not to build. Artificial intelligence is changing the distance between wanting software and making it.

AI & Edtech · August 13, 2026Edtech Is Still Mailing DVDsHow 90% of edtech disappears: people have not stopped learning, they have started consuming knowledge differently, and that distinction may wipe out much of the industry.

← Back to Blog

---

### https://sahra-josephine.com/blog/edtech-customer-competitor

← Blog

AI & Edtech·September 5, 2026When Your Customer Becomes Your Competitor: Your Customer Found a Shortcut to the Software Factory

Part four of how 90% of edtech disappears. Software companies spent twenty years convincing customers not to build. Artificial intelligence is changing the distance between wanting software and making it.

I spent part of my summer holiday walking through one of the most beautiful commercial failures ever built. Park Güell is now prime Barcelona: Gaudí, mosaics, extraordinary views and a reliable supply of tourists photographing themselves beside a ceramic lizard. But it was not intended to be a public park. Eusebi Güell and Antoni Gaudí planned it as an exclusive residential estate for wealthy families, with sixty homes set above the city. Only two were ever built.

Our official guide explained that one of the problems was surprisingly practical. The people wealthy enough to live there still needed to reach their factories and the port, which meant travelling by horse-drawn carriage along an inconvenient route, with the journey affected by weather and road conditions. The estate may have offered cleaner air and beautiful views, but it was simply too difficult to get from the house to the place where the money was made.

There were other obstacles. The plots came with restrictive conditions, public transport was inadequate, and the exclusivity that made the development attractive also helped make it commercially unviable. Construction stopped in 1914. Güell's heirs eventually sold the land to the city, and it opened as a public park in 1926.

Today, the location no longer feels remote. Barcelona grew around it, while roads, buses, taxis and the metro changed the practical meaning of distance. Park Güell was not necessarily built in the wrong place. It was built before the road arrived.

I have spent eleven years in SaaS in the edtech industry telling customers not to build software they could buy. Ninety-nine percent of the time, this was genuinely excellent advice. Building your own learning platform meant developers, designers, product managers, infrastructure, integrations, security, maintenance and enough budget to survive the point eighteen months later when somebody discovered that the internal system everyone had worked so hard on was worse than the product they could have bought in the first place.

The software factory existed, but for most companies reaching it was slow, difficult and prohibitively expensive. So we taught companies to rent. Do not build your own learning management system. Buy one. Do not build an authoring tool. Subscribe to one. Do not build an employee academy or assessment platform. Find a specialist company that has already solved the problem. I built a successful, venture-backed edtech company on that logic.

Today, I also own an AI studio, and I increasingly give customers what appears to be the opposite advice: before buying another software product, find out what it would cost to build the functionality you actually need. That does not necessarily mean building it in-house. A company can ask one of its own developers, an AI-assisted contractor or a studio like mine. It does not need to become a software company or own the factory. It simply needs affordable access to one.

I know how disruptive that advice is because I have followed it. I have previously written about my personal newsletter costing approximately 10,000 Danish kroner every month to send emails to around 36,000 people through Mailchimp. I used Lovable to build my own sending tool and connected it to SendGrid. Building it cost less than $30, operating it costs approximately $25 to $30 a month, and the part I did not mention previously is that it took less than a day.

The tool is still not as good as Mailchimp and lacks some of its integrations, templates and edge cases. But I don't need them. It performs the fraction of Mailchimp's functionality I need, replacing a subscription costing roughly 10,000 Danish kroner ($1,520) a month, or around $18,000 a year, with infrastructure costing approximately $25 to $30 a month, or $300 to $360 a year. Mailchimp did not lose me to another email platform. It lost me to myself.

AI is paving the road between the company office and the software factory, which is why I increasingly catch myself asking customers a question I would once have considered terrible advice: Why are you buying this at all?

The purchase used to exist before competition began

Most software companies think about competition after the customer has decided to buy. The customer needs a learning platform, so it compares learning platforms. It needs a CRM, so it compares CRMs. It needs a newsletter system, so Mailchimp competes with HubSpot, Klaviyo and whatever else appears in the shortlist. The category has already won; the only remaining question is which vendor receives the money.

That assumption sits underneath an enormous amount of SaaS strategy. Marketing generates demand for the category, sales converts that demand into a named account, product adds enough features to beat the nearest competitor, and customer success protects the renewal. Even churn presupposes that a customer existed first.

Take a company with 10,000 employees and a terrible learning setup. Ten years ago, HR might have invited five LMS vendors to demonstrate their products and selected one. Now ask what the organisation actually needs. It already has an identity system, policies, videos and internal knowledge. AI can turn that material into exercises. What remains is storing completion, reporting to managers, documenting compliance and connecting the pieces.

Historically, even that narrow workflow required enough engineering effort that buying the platform remained rational. Now the company can describe the workflow, connect its existing systems and have an internal product person, an AI-assisted developer or an external studio build the part it needs. It can still buy if the commercial product is better, but buying is no longer the automatic starting point. The company does not have to recreate the LMS. It has to recreate the reason it bought the LMS.

The most dangerous prospect is therefore not the customer who leaves at renewal. At least that customer entered the CRM, signed a contract and generated a reason for leaving that somebody can analyse. The more unsettling prospect is the company that never becomes a prospect: it never visits the pricing page, appears in the pipeline, requests a demonstration or gives procurement a shortlist. Somebody asks whether this needs to be another subscription, discovers that it does not, and takes the new road to the software factory instead. From the SaaS provider's point of view, no deal was lost because no deal ever existed.

You only have to beat the invoice

For most of the SaaS era, vendors could bundle hundreds of capabilities because reproducing the subset a customer used was expensive enough to make renting the whole product sensible. That threshold is falling. The customer does not have to build something better, recreate the vendor's entire platform or cover every edge case. It only has to solve its own problem well enough to beat the invoice.

Take a deliberately simple learning-platform price of $12 per user per month, with no implementation fee, base fee or volume discount. A company with 10,000 users pays $1.44 million a year. At 50,000 users, the annual cost is $7.2 million; at 200,000, it is $28.8 million. This is not a claim that a competent procurement team would accept that flat rate at 200,000 seats. Even after a 75 percent volume discount, however, the annual subscription would still be $7.2 million. The inputs change the point at which building becomes rational; they do not remove the difference between a cost that scales per seat and one that may not.

Of course, a customer of that size would negotiate. Enterprise contracts are rarely that clean, and a serious learning platform does far more than host a few pages and record completion. It may provide hundreds of integrations, sophisticated permissions, accessibility, audit trails, content standards, localization, support, security documentation and contractual accountability. A credible comparison has to include those things.

But that is precisely the point: the customer does not have to reproduce everything the vendor has built. Imagine, purely as a model, that a tailored learning platform costs $1 million to build and another $2 per user per month to operate, secure and maintain. For 10,000 users, the first year would cost approximately $1.24 million, already below the $1.44 million subscription. For 50,000 users, it would cost approximately $2.2 million rather than $7.2 million; for 200,000, approximately $5.8 million rather than $28.8 million.

Those are illustrative numbers, not a universal business case for building. The calculation can reverse quickly when a company needs global compliance, dozens of deep integrations, round-the-clock support, complex migrations or a supplier willing to assume real operational liability. Custom software creates maintenance, security and technical-debt risks that a spreadsheet can conceal with impressive efficiency. Many enterprise platforms also price by active users, usage or negotiated bands rather than applying a flat public rate.

Yet even after generous caveats, the direction is difficult to ignore. Per-user SaaS pricing rises with the size of the customer, while the cost of building the relevant functionality does not necessarily rise at anything close to the same rate. Larger customers mean more seats and expansion revenue, but they may also have the strongest financial reason to question whether the vendor should exist in the workflow.

Nobody will spend six months recreating a product that costs €200 a month. A small school should probably not maintain a homemade student-information system, and a regulated organization should be careful with sensitive data. A company paying €250,000 a year has a different calculation. Your most attractive customer may also have the strongest incentive never to become your customer.

The old question was whether a custom build could equal the product. The new question is whether it can beat the invoice, which is a much lower bar.

The rent moves down the stack

This is not the end of renting; it is a change in what the customer rents. My Mailchimp replacement still uses SendGrid to deliver email. I did not build global email infrastructure, negotiate directly with every internet provider or create my own system for protecting sender reputation. I replaced an application subscription with a thinner collection of infrastructure and a small piece of software I control.

A company building its own learning environment will probably do the same. It may pay OpenAI, Anthropic or Google for model access, use AWS or Azure for infrastructure, and retain an external team to maintain the system. The software factory has not become free. It still needs electricity, machinery, raw materials and somebody who understands what is being made, but the rent moves down the stack.

Instead of paying a large recurring fee for a finished application containing 400 functions, the customer can pay smaller recurring fees for infrastructure and own the seven functions it actually uses. It can commission a one-time build without creating an in-house software department, just as a fashion company can commission production without owning every machine that makes its clothes. The choice is no longer simply build or buy. It is buy, build, assemble or commission, and AI is reducing the cost of the final three.

SaaS originally won because one specialist company could build a product once and distribute it cheaply to thousands of customers. That advantage does not disappear, but the vendor must recover the cost of a general product, its sales organisation, customer-success team, investors and feature roadmap across its customer base. A custom system only has to serve one company. For years, that was also its weakness because one customer could not justify the journey to the factory. Now the road is shorter.

Edtech is particularly exposed because two costs are falling at once. Generative AI reduces the cost of producing and coordinating learning: a manager can ask an assistant to explain a policy, generate an exercise, adapt material to a role, translate it and test understanding without opening a course catalogue. At the same time, AI-assisted development reduces the cost of building the system that assigns, tracks and documents that learning. One force attacks usage; the other attacks the purchase.

An edtech company can therefore be squeezed even while demand for learning grows. Organizations will still train employees, distribute knowledge, document compliance and develop skills, and may do more of all four. But increasing demand for the outcome does not guarantee increasing demand for the existing product category. People did not stop watching films when DVD rental collapsed, or listening to music when buying CDs became absurd. The activity survived; the product, distribution and payment model around it changed.

Learning is not disappearing. The assumption that it must be packaged inside a separately purchased learning platform is becoming less secure.

Maybe your moat was the distance

The obvious response is that vibe-coded software is unreliable, insecure and easy to demonstrate but difficult to operate. Often, that is true. There is an enormous distance between making a functioning prototype and running a business-critical system, which needs architecture, tests, permissions, monitoring, backups, data governance, accessibility, security reviews and somebody accountable when it breaks. AI can produce bad code very quickly and allow a company to create a fragile internal dependency that nobody understands six months later.

But weak execution does not rescue a weak commercial model; it merely means customers need a competent route to building. The early internet was full of terrible websites, and that did not protect newspaper classifieds, travel agents or high-street retailers. Poor first attempts can coexist with a structural change in cost and access.

The more useful question is which part of a SaaS company's defensibility came from genuinely difficult work and which part came from the customer being too far from the factory. Deep integrations, proprietary data, regulatory approval, contractual liability, trusted credentials, distribution, community and demonstrated outcomes can be formidable. A supplier that understands a complicated domain and accepts responsibility for running it can be worth far more than its features.

Having many features is not necessarily a moat. A beautiful interface is less durable when interfaces can be generated, a decade of accumulated code matters less when the customer needs only a narrow workflow, and switching costs offer limited protection against a company that has not bought anything yet. Many SaaS businesses were safe partly because recreating even an inferior version required a team the customer did not possess and a budget it could not justify.

AI does not make every product easy, safe or sensible to rebuild. It makes enough products cheap enough to question. A harsher definition of defensibility follows: a moat is not what makes your product difficult to recreate. It is what makes the customer's purchase difficult to eliminate.

For twenty years, software positioning mostly answered why us instead of them? The customer had already accepted that it needed to buy something, and the vendor's job was to win the comparison. Increasingly, software companies will have to answer a more dangerous question: why buy at all?

That is a much harder argument because the company is not objecting to your price, requesting another feature or threatening to select your competitor. It may never enter your market in the first place. No opportunity appears in the CRM, no procurement process begins and no lost-deal analysis explains what happened. The need is simply resolved through software the company owns rather than software it rents.

The functionality remains. Employees still learn, managers still need information and compliance still needs evidence. What disappears is the transaction in the middle.

Park Güell was not built in the wrong place. It was built before infrastructure changed what that place meant. SaaS was built for a world in which the road to the software factory was too long and expensive for most companies to travel, so it turned the factory's output into something they could rent. That was genuinely good advice until the road changed.

The software factory has not disappeared, and neither has the need for what it makes. But companies can increasingly reach it without passing through the SaaS provider that expected to sell them a subscription, meaning that your competitor is not another product. It is the disappearance of the purchase.

Series · How 90% of Edtech Disappears

A four-part series on the money, the market, and the models that ended an industry as we knew it.

Part 1 of 4If I Operated an Edtech Fund, I Would Be Shitting My Pants

Part 2 of 4Your Edtech Investor Wants an Open Relationship

Part 3 of 4 · Previous in seriesEdtech Is Still Mailing DVDs

Part 4 of 4When Your Customer Becomes Your Competitor: Your Customer Found a Shortcut to the Software Factory

Read the whole series →Related essays

AI & Edtech · July 18, 2026If I See One More Person Push “PedTech,” I’m Going to VomitPedagogy was never the missing idea. The harder problem was building an industry whose economics allowed pedagogy to stay at the center once investment dollars came knocking.

AI & Society · September 8, 2026Elon Musk Bought the Wrong Real Estate for the Everything AppMusk bought Twitter to build the everything app. But while he was accumulating attention, OpenAI was accumulating intent, and that may be the most valuable address on the internet.

AI & Society · August 16, 2026AI Ate the Internet. Now We Want It to Decide What’s Human.Watermarking is supposed to make synthetic content transparent. It may also turn AI companies into the institutions we ask to certify human authorship.

← Back to Blog

---

### https://sahra-josephine.com/blog/ai-human

← Blog

AI & Society·August 16, 2026AI Ate the Internet. Now We Want It to Decide What’s Human.

Watermarking is supposed to make synthetic content transparent. It may also turn AI companies into the institutions we ask to certify human authorship.

When I was a kid, I wrote my school essays by hand and then typed them out on our typewriter.

I was supposed to do the typing myself. It was useful to learn. But sometimes it got late, the essay was already written, I was exhausted, and my mother would take over so I could go to bed. She was a single mother with better things to do than supervise an overtired child ceremonially pressing every remaining key at 11 p.m.

Nobody thought she had therefore written my essay. Nobody called it plagiarism, cheating, or a fake assignment. We understood intuitively that there was a difference between the capability being assessed and the infrastructure used to produce the final artifact. The ideas, argument, language and understanding were mine. Typing had some value as a skill, but it was not the point of the essay.

Then we got a computer. This was extremely exciting, not least because the same machine on which I could suddenly write my essays also let me play Gorillas, the wonderfully stupid game where two gorillas threw exploding bananas across rooftops.

Now the computer provided the typing infrastructure directly. My mother became the spellchecker, reader, and person I bounced ideas off. Then software became the spellchecker. Grammar software started correcting sentences. Google became part of the research process. Autocomplete began predicting what I was trying to write.

Nobody demanded a receipt documenting exactly which words had been touched by my mother, Microsoft Word, or Google.

And then AI arrived.

Why did we decide this is where infrastructure becomes authorship?

I use AI for some combination of all those previous roles: typist, researcher, editor, critic, spellchecker, and brainstorming partner. Some of that is infrastructure. Some is intellectual assistance. Sometimes it is genuine intellectual labor.

But intellectual labor and authorship have never been synonyms. My mother contributed intellectually when I bounced ideas off her. An editor can substantially improve an argument. A researcher can discover the fact on which an essay depends. None automatically becomes the author.

Authorship is closer to where the governing intellectual agency sits: who decides what the work is about, chooses between ideas, understands the argument, determines what belongs in it, and takes responsibility for the result.

Years ago, we experimented at CanopyLAB with using artificial intelligence to assess student work in almost the opposite way from what watermarking does today. A single grade was too crude. We wanted to distinguish between things like understanding, argument strength, coherence, grammar, and style, then use those distinctions to give better qualitative feedback.

At the time, the technology made that difficult. Then generative AI arrived, and much of the technical problem became almost irrelevant. Modern models can look at a piece of writing and separate dimensions of performance in seconds. So I do not believe the problem is that we could never know what Claude contributed. We probably could know much more.

Anthropic says future Claude models will watermark text as part of its response to EU transparency requirements. Today, that watermark can indicate that Claude was involved without necessarily distinguishing between Claude generating something and Claude heavily editing it. But there is no reason to assume provenance has to remain that crude forever.

Perhaps the future receipt is itemized. Claude proposed the structure. Claude found the counterargument. Claude rewrote four paragraphs. Claude corrected the grammar. The human rejected six suggestions, verified the sources, rewrote half the generated text, and made the final decisions.

Fine. We now have the full ingredient list. We still have not answered who cooked the meal. That is the distinction I care about. Better measurement of contribution does not automatically give us a theory of authorship.

The Ministry of Authenticity

Anthropic does not claim to decide who authored a piece of work. The more interesting reality is that everyone else seems to cast it in that role.

A university wants to know whether a student used Claude. A publisher wants to enforce an AI policy. An employer wants to check an application. A literary prize decides how much AI assistance it will tolerate. Soon there is an API capable of answering a question those institutions cannot answer themselves: was Claude likely here?

Nobody at Anthropic has to announce a Ministry of Authenticity. Yet we are building one around them ourselves. And this is where the consequences become more interesting than whether the watermark works.

Imagine a student writes an essay, uses Claude to challenge the argument and tighten the language, and the watermark correctly identifies Claude’s involvement. The detector has not made a mistake. But a school can still make one if it treats that accurate signal as proof that the student did not do the intellectual work.

The same could happen to a writer who uses AI as an editor, a job applicant who uses it to improve awkward phrasing, or a researcher writing in a second language who uses it to polish a grant application.

The most dangerous version of this system is not one that falsely detects AI. It is one that correctly detects AI participation and then draws a completely unjustified conclusion about the human.

And incentives do what incentives do. If possessing the receipt becomes punishable, people will learn to destroy the receipt. The people using AI openly and legitimately become the easiest to identify, while those actually trying to conceal its role acquire every reason to remove the watermark, “humanize” the text, or move to tools that leave no trace.

We may therefore introduce provenance to create transparency and end up incentivizing concealment instead. The technology can work exactly as intended. The problem begins when we reward and punish people based on what we decide the receipt means.

Infrastructure is not competence

Education makes the distinction obvious. If I want to test mental arithmetic, take away the calculator. If I want to test typing, my mother finishing my typing defeats the purpose. But if I want to know whether somebody can analyze evidence, construct an argument, and defend a conclusion, assess that.

Ask where the model was wrong. Ask what the student accepted and rejected. Ask why they chose one argument over another. Ask them to defend what they submitted and take responsibility for it.

“Did AI touch this?” is easier to measure. It is also a different question.

Before AI, the finished artifact was a reasonably useful proxy for the mind that produced it. A student handed in an essay, and we inferred something about their understanding from the essay. AI breaks that proxy. But the answer does not have to be rebuilding it through increasingly sophisticated forensic analysis of the artifact. We can assess the mind more directly.

And we already know we can do better. The whole promise of using AI in assessment was that we could move beyond crude signals and understand more about the human: what they understood, where their reasoning was strong, what they struggled with, and what they should work on next.

The point was empowerment through richer feedback. It would be a strange reversal if considerably more capable AI now helped us reduce that same human performance to a forensic label about which tool participated.

Tools exist because opportunity cost exists. I could calculate a spreadsheet manually, but Excel exists. I could search twenty books for a statistic I can now find in minutes. I could spend three hours mechanically cutting repetition from an article, or spend some of those hours developing the next argument.

The fact that I could perform the lower-value task myself does not make performing it myself intellectually superior.

Much technological progress consists of turning expensive or time-consuming tasks into infrastructure and freeing human attention for something else. AI may be the next layer of intellectual infrastructure.

That does not mean understanding becomes less important. I think the opposite may be true. The more execution we can outsource, the more important judgment becomes. If a machine can produce a plausible answer in seconds, knowing whether the answer is good, bad, fabricated, derivative, or irrelevant becomes considerably more valuable.

The skill moves. Our assessment should move with it.

And somehow the machine gets the better receipt

Generative AI became useful by learning from an enormous environment of human intellectual production. Writers wrote. Researchers researched. Programmers coded. Teachers explained. Millions of people created the information environment from which these systems became valuable.

There are unresolved legal questions about training data, and that is not my argument here. I am interested in the direction of attribution.

Human knowledge flows into AI systems at enormous scale, and individual contribution becomes diffuse. Then the machine participates in our next piece of work, and suddenly its contribution becomes important enough to make technically traceable.

Humanity supplies the intellectual raw material. The machine gets the receipt.

Make that receipt as perfect as you like. Tell me exactly which sentence Claude proposed, which source it found, which paragraph it restructured, and which comma it moved. I still do not think machine provenance should become the proxy by which someone else decides whether the work is mine.

We understood the distinction when an exhausted child went to bed, and her mother finished typing an essay that was already written. Nobody needed a forensic record of how many keys my mother pressed or which spelling corrections came from her. We used judgment because we understood what the assignment was actually supposed to measure.

Perhaps we should remember that before building an authenticity infrastructure around the easiest thing to detect.

The receipt tells us the machine was there. It still cannot tell us who thought.

Related essays

AI & Society · September 8, 2026Elon Musk Bought the Wrong Real Estate for the Everything AppMusk bought Twitter to build the everything app. But while he was accumulating attention, OpenAI was accumulating intent, and that may be the most valuable address on the internet.

AI & Edtech · September 5, 2026When Your Customer Becomes Your Competitor: Your Customer Found a Shortcut to the Software FactoryPart four of how 90% of edtech disappears. Software companies spent twenty years convincing customers not to build. Artificial intelligence is changing the distance between wanting software and making it.

AI & Edtech · August 13, 2026Edtech Is Still Mailing DVDsHow 90% of edtech disappears: people have not stopped learning, they have started consuming knowledge differently, and that distinction may wipe out much of the industry.

← Back to BlogAbout Sahra-JosephineSpeakingInvite me to speak

---

### https://sahra-josephine.com/blog/edtech-dvds

← Blog

AI & Edtech·August 13, 2026Edtech Is Still Mailing DVDs

How 90% of edtech disappears. People have not stopped learning. They have started consuming knowledge differently. That distinction may wipe out much of the edtech industry as we know it.

When I went to college, Netflix was a company that sent DVDs through the mail. You made a list of everything you wanted to watch, which Netflix called your “queue”. Its standard subscription plan in 2002 allowed you to have three DVDs out at the same time, with no due dates or late fees, and the company also offered other plans with different service levels. When you returned one, Netflix sent another available title from your queue.

At the time, it was brilliant. You did not need to drive to Blockbuster. There were no late fees. The catalogue was enormous compared with what a physical store could stock, and Netflix could learn what you liked and recommend what you should watch next. But there was one enormous limitation Netflix could not optimise its way out of: you still had to wait for the next DVD to arrive.

When Netflix launched streaming in 2007, it began removing the physical delivery altogether and subsequently expanded internet delivery beyond the PC to other devices. The customer relationship remained, the recommendation system remained, and much of what created value remained. The red delivery envelope and the wait did not.

That distinction is increasingly how I think about edtech. I have argued in my first essay that as much as 90% of today's edtech will disappear, consolidate, become irrelevant or be absorbed into something else over the next few years. I do not believe that because education is becoming less important. I believe it because learning is having its Netflix moment.

People are still searching for information. They are still trying to understand difficult concepts, practising skills, preparing for exams, learning languages, changing careers, solving problems and asking for explanations. What has changed dramatically is how they do it.

They open ChatGPT, Claude, Gemini or Perplexity. They ask the question directly and then ask another one. They upload the document they are struggling with. They ask for an explanation at the level of a 12-year-old, then at the level of a graduate student. They ask for examples and practice questions, argue with the answer, ask the system to quiz them and ask it to explain where they went wrong.

This is no longer hypothetical consumer behaviour. OpenAI reported in 2025 that more than one-third of college-aged young adults in the United States used ChatGPT and that approximately one-quarter of their messages were related to learning or schoolwork. UNESCO reported that more than two-thirds of secondary-school pupils in high-income countries were already using generative AI, while Anthropic's 2026 usage data placed educational instruction and library work as the second-largest category of Claude.ai usage.

The learning did not disappear. The mode of consumption changed. And much of edtech is still mailing DVDs.

The uncomfortable history of edtech

The obvious explanation for why so much of edtech now looks vulnerable is generative AI. It would be convenient to say that ChatGPT arrived and suddenly made a healthy industry obsolete. Unfortunately, the chronology does not support that explanation.

The industry had serious structural problems long before ChatGPT appeared. If you look at the past 10 or 15 years of edtech, you can almost see a succession of attempts to solve the problems created by the previous generation. The early enterprise systems were often horrible. They were difficult to buy, difficult to configure and difficult to use. Implementing one could require consultants, integrations, training, migration projects and months of work before a learner saw anything at all.

Then SaaS arrived and appeared to solve the problem. Products became easier to buy, easier to implement and better designed. They were hosted in the cloud and subscription-based, with no giant IT project required. This was genuine progress, but somewhere along the way the industry began confusing the removal of implementation friction with the creation of value.

We became extraordinarily good at making software: a course platform, a quiz platform, a tutoring platform, a teacher platform, an engagement platform, a language app, a wellbeing app, a financial-learning app, a professional-development app, a skills platform. Another login, another dashboard and another subscription. Because SaaS made software easier to sell, an enormous amount of software was sold. That did not mean anyone used it.

UNESCO's 2023 Global Education Monitoring Report found that an average of 67% of education-software licences in the United States were unused and 98% were not used intensively. The EdTech Genome Project cited in the report looked at roughly 7,000 pedagogical tools representing $13 billion of spending and found that 85% were either a poor fit or implemented incorrectly. That is an astonishing number, and I think the industry has been too comfortable calling this an implementation problem.

Of course implementation matters. Schools fail to train teachers. Leaders buy software without changing behaviour. Products integrate badly. Teachers do not have time. Procurement and usage are disconnected. But from the perspective of a business model, those explanations only get you so far.

If I build a product that produces value only when the customer trains everyone correctly, integrates it perfectly, continually champions usage and somehow persuades thousands of busy humans to develop a new habit, those dependencies are part of my product economics. I cannot simply declare them to be the customer's problem.

The evidence problem was just as uncomfortable. UNESCO found that only 7% of the UK edtech companies it reviewed had conducted a randomized controlled trial. Twelve per cent had used third-party certification, and 18% had participated in academic studies. In a survey across 17 American states, only 11% of teachers and administrators requested peer-reviewed evidence before adopting education technology.

So we created a peculiar market. Companies frequently did not prove that their products improved learning, buyers frequently did not demand proof before purchasing them, and the industry became extremely sophisticated at measuring whatever software happened to make measurable: registrations, logins, minutes watched, questions answered, courses completed and certificates generated.

Those things tell you what happened inside the software. They do not necessarily tell you that somebody learned, and they certainly do not tell you that the customer received enough value to keep paying.

COVID temporarily made almost every DVD look useful

Then COVID arrived. Global edtech venture funding was approximately $7 billion in 2019. It jumped to $16.1 billion in 2020 and $20.8 billion in 2021. By 2024 it had collapsed to approximately $2.4 billion, before moving only slightly upward to $2.6 billion in 2025.

The usual interpretation is that investors went insane during the pandemic and then came back to their senses. There is some truth in that, but COVID also created an extraordinary product-market-fit illusion because it removed the alternative.

Schools did not voluntarily select online education over fully functioning classrooms. Universities could not wait until the software improved. Employers could not simply cancel every training program indefinitely, and parents could not decide that their children had enough screen time and send them back to school. There was nowhere else to go.

I experienced this directly. Before the pandemic, companies like mine could spend the first half of a sales conversation explaining why learning should happen online in the first place. Suddenly we did not have to. The world had made the argument for us.

Usage exploded, but emergency adoption proves that a product can function when people have no alternative. It does not prove that they will continue choosing it when the alternative returns. The industry treated emergency traffic as though an entirely new behavioral pattern had been permanently established. Then the world reopened, and customers got their choice back.

If COVID temporarily disguised structural weaknesses that were already there, those weaknesses should have started showing up before generative AI became a serious substitute. The obvious place to look for that evidence is the public market.

The market broke before ChatGPT

I wanted to know whether the weaknesses that COVID had temporarily hidden had started showing up before generative AI became a serious substitute. Not because share prices can tell us whether people wanted to learn, they cannot, but because they can tell us something much narrower and still very important: what happened to the money investors put into the companies that were supposed to benefit from the digitisation of education.

So I went back to the end of 2020 and took the 20 largest pure-play listed education companies by market value at the time. They were TAL Education Group, Offcn Education, New Oriental, Chegg, Gaotu Techedu, Bright Horizons, Pearson, Kahoot!, Grand Canyon Education, China Education Group, China East Education, IDP Education, East Buy Holding, China Yuhua Education, Youdao, Laureate Education, Pluralsight, 2U, Cornerstone OnDemand and John Wiley & Sons.

Then I ran a deliberately simple experiment. Imagine that on 31 December 2020 you had $1 million and decided that edtech was where you wanted to put it. What would have happened to that money?

There are obvious limitations to what this exercise can prove. A share price is not a measure of pedagogy, learner demand or product quality. A company can grow revenue while destroying shareholder value because investors originally paid too much for it. Another can build an excellent product and still produce a mediocre return. Share prices are affected by interest rates, execution, competition, regulation, geography and the expectations already embedded in the starting valuation.

So I am not using the stock market as evidence that people stopped wanting education. I am using it to ask a much narrower question: were investors actually rewarded for allocating capital to the companies that represented the category?

There was also one enormous complication in the original 20. Nine of the companies were Chinese, and China's 2021 tutoring restrictions fundamentally changed the economics of large parts of the private education market. Those losses are completely real if you were an investor, but they are less useful if the question is whether the underlying business models were already weakening independently of an extraordinary regulatory event. So I calculated both.

The full 20-company portfolio tells us what actually happened if an investor bought the category as it existed at the end of 2020, including China. For the central comparison, I then removed the nine Chinese companies and looked at the remaining 11: Chegg, Bright Horizons, Pearson, Kahoot!, Grand Canyon Education, IDP Education, Laureate Education, Pluralsight, 2U, Cornerstone OnDemand and John Wiley & Sons.

Then I asked what happened to the $1 million.

In the ex-China group, $1 million invested in the median company became approximately $1.07 million. If the money had been divided equally across all 11 companies, it would have become approximately $1.31 million, largely because a small number of strong performers pulled the portfolio upward. If instead the $1 million had been allocated according to the companies' market values at the beginning of the period, effectively following the hierarchy investors had already assigned to the category, it would have become approximately $1.02 million.

Over the same period, $1 million invested in SPY became approximately $2.10 million. In QQQ, it became approximately $2.18 million. That comparison is difficult to make flattering.

The specialist edtech investor had to understand school procurement, pedagogy, implementation, course completion, regulation, institutional budgets, B2B sales cycles and the difference between an LMS, an OPM, a tutoring company and a university operator. They had to decide which companies deserved capital, which valuations were justified and which business models would survive. The passive investor could buy SPY and go to lunch.

Following the original market hierarchy of the ex-China education companies left the specialist investor with approximately $1.02 million. The same starting amount in SPY produced approximately $2.10 million. More than $1 million in additional value came from choosing the passive alternative.

That does not prove that edtech was useless, nor does it prove that there was no demand for digital education. It proves something much more uncomfortable for an investor: as a group, the largest listed education companies did not reward the additional specialization, concentration and risk required to invest in them.

But even that conclusion is too broad, because the companies did not perform remotely alike.

One million dollars invested in Laureate Education became approximately $5.67 million. In Pearson, approximately $2.21 million. Grand Canyon Education became approximately $1.71 million, Wiley approximately $1.45 million and Cornerstone OnDemand approximately $1.31 million.

At the other end, $1 million invested in Chegg fell to approximately $12,000. In 2U, it went to zero. IDP Education fell to approximately $122,000, Kahoot! to approximately $368,000 and Bright Horizons to approximately $478,000.

Those are not minor differences around a common sector result. One company multiplied the original investment more than five times while another effectively erased it.

That matters because it rules out the simplest conclusion. The market was not deciding that education itself had become worthless. It was distinguishing between very different kinds of businesses that happened to have been placed under the same edtech label.

In fact, that may be part of the problem with the category itself. We call an enormous range of fundamentally different businesses “edtech” because they all touch education. But a university operator, a publishing company, an assessment provider, an LMS, a tutoring platform and a consumer learning app do not necessarily share the same economics simply because their customers are learning something.

Some of the stronger businesses were attached to institutions, credentials, assessment, research, publishing, regulated operations or infrastructure deeply embedded inside customers. Others depended much more heavily on people repeatedly choosing to visit a separate digital destination for content, courses, activities or answers. And that second group contained some of the most spectacular destruction.

Which creates the obvious next question. If those businesses were especially vulnerable because users had to keep coming back to a separate platform for something they could increasingly obtain elsewhere, was ChatGPT the thing that broke them?

The chronology says no.

The first collapse happened before AI. Then came a second one.

Chegg is the most violent example. At its peak in 2021, the company was worth approximately $14.7 billion. Today its market value is approximately $100 million. More than 99% of the value disappeared.

It is extremely tempting to look at that destruction and tell a simple story: ChatGPT arrived, students stopped paying for answers, and an entire category broke. The problem is that the first break happened too early.

I separated the listed education companies according to whether general-purpose AI could reproduce a central outcome customers paid them to provide. This was not a classification of whether the companies themselves happened to use AI. It was a measure of how easily a general-purpose AI system could substitute for the outcome the customer had previously needed that company to obtain.

From February 2020 through November 2022, $1 million invested in the median higher-exposure company fell to approximately $550,000. Over the same period, $1 million invested in the median lower-exposure education business became approximately $1.39 million.

That happened before ChatGPT had any meaningful opportunity to cause the collapse. Something was already wrong.

Then the pattern did something important that is easy to miss if you simply compare the beginning and the end: it stabilised. From November 2022 through December 2023, $1 million invested in the median higher-exposure company became approximately $1.04 million, while the equivalent investment in the lower-exposure group became approximately $1.06 million. For roughly a year, the gap stopped widening. The vulnerable group had already been severely repriced, but it was no longer collapsing relative to the rest of the sector.

Then came the second break.

During 2024 and 2025, as generative-AI adoption accelerated, $1 million invested in the median higher-exposure company fell to approximately $400,000. The lower-exposure equivalent became approximately $1.08 million.

That sequence is much more interesting than saying that AI killed edtech. First, the businesses most exposed to substitution fell sharply before ChatGPT. Then they stabilized. Then, once consumers had a radically easier way to obtain many of the same outcomes, the gap opened again.

That does not prove that AI caused every dollar of the second decline. Share prices are affected by interest rates, execution, competition, geography and starting valuation. But it establishes the order of events, and that order makes the comfortable explanation much harder to sustain.

If generative AI had suddenly broken an otherwise healthy industry, the first collapse should have followed the arrival of generative AI. It did not.

The more uncomfortable conclusion is that a meaningful part of edtech had already developed weak economics around products that were difficult to retain, poorly utilized or dependent on users repeatedly visiting separate destinations. The first market correction exposed those weaknesses. For a brief period, the damage stabilized. Then generative AI changed the way people could access many of the same outcomes, and the companies whose value propositions were easiest to reproduce started falling again.

That distinction matters enormously. If ChatGPT were the original problem, the obvious response might be to add AI to the existing product. But if the underlying need remains while the way people access it is changing, then the strategic question is much bigger: does the same value still need to be delivered through the same product at all?

What if the product is useful but the container is wrong?

Netflix became successful not because people suddenly wanted different movies, but because new technology allowed it to change the delivery mechanism and remove friction for the user. Today, the equivalent friction is rarely a physical barrier. It might be a payment model, a separate login, an app somebody has to remember to open, a course they have to commit to completing or a platform that requires people to change an existing habit.

Learning companies are now competing against learning that happens through TikTok, podcasts, YouTube and general-purpose AI. In a 2025 European study commissioned by YouTube, 74% of the young people surveyed said they had watched YouTube videos to learn something new for school. Sesame Street's official YouTube channels generated more than five billion views in the year before YouTube announced an expanded partnership that would make the platform the largest digital home for Sesame Street episodes from 2026.

Again, the demand for learning did not disappear. It moved. Picking the wrong container can therefore result in dramatic failure. A recent Danish example makes the distinction unusually concrete. MYiNNERME was a digital mental-health app developed by child psychologists for children, teenagers and their parents. It offered self-guided learning programs addressing anxiety, anger, shyness, grief, bullying, sleep and self-esteem, translating established psychological approaches into exercises and content families could use outside a traditional therapy session. The company behind MYiNNERME has now closed down, citing the inability to find a profitable business model.

The obvious conclusion would be that the product failed because there was insufficient demand for what it offered. I am not sure that is the right conclusion.

If I were approaching the same problem today, I would question whether an app should be the starting point at all. Imagine instead a faceless YouTube channel built around the company's characters. A child might encounter one of those characters in a video explaining anxiety, meet the same character in a story about bullying, use it for a breathing exercise and later recognize it in content about going to sleep or starting a new school.

Creating that universe no longer requires anything close to the production infrastructure it once would have. Characters can be created and animated with AI-assisted tools. Scripts, voices, translations and variations can be produced far more cheaply. Production and publishing can be automated. The same characters and psychological concepts can move from long-form YouTube into shorts, audio, printable exercises and material used by psychologists or schools.

They can also move outside the screen entirely. The same intellectual property could become stories, print-on-demand coloring books, activity books, cards or other physical products without the company needing to finance enormous inventories before discovering whether anyone wants them.

The point is not that MYiNNERME would necessarily have succeeded as a YouTube company. It is that the valuable asset was never necessarily the app. It was the expertise, the psychological frameworks, the exercises, the stories, the characters and the ability to translate difficult emotional concepts into something children could understand. An app was one way of wrapping that value. Today there are many more.

One particular company and one particular delivery model can therefore disappear while the underlying demand remains completely intact. In fact, that demand may now be served in more places than ever before: inside dedicated products, through creators, on YouTube, through AI, in schools, between therapy sessions and through physical products connected to digital worlds.

But this creates another problem. If software becomes cheaper to build and content becomes cheaper to create, then simply owning good content may not be enough either.

The important question becomes: what remains scarce? It may be trusted brands. It may be credentials. It may be proprietary data, distribution, institutional relationships, community, regulatory approval, workflow integration, recognized assessment, intellectual property or demonstrated outcomes. The answer will differ by company, but the principle is the same. When both software production and content production become cheaper, investors have to identify the part of the business that cannot easily be reproduced. That is a much more useful distinction than asking whether an edtech company “uses AI”.

Edtech's DVD moment is also an investor problem

Generative AI has not only created new places for people to learn. It has also made it dramatically cheaper for companies to repackage the same intellectual property across many of those places.

But there is an important difference between MYiNNERME, the broader edtech industry and Netflix. Edtech is not entering this technological transition from a position of obvious economic strength. The financial problems and declines in value predate generative AI. Long before ChatGPT became a serious substitute, the sector had unused licenses, weak evidence of outcomes, difficult retention and business models that depended on users repeatedly returning to separate digital destinations. The public markets were already reflecting some of those weaknesses.

That is why I increasingly think the useful distinction is not simply between “AI companies” and “non-AI companies”. It may be between destinations and infrastructure. A destination requires the customer to decide to come to you. Open the app. Visit the platform. Start the course. Watch the lesson. Ask the tutor. Complete the activity. Return tomorrow.

Infrastructure is different. It sits inside a workflow, institution or system that already exists. Credentials, assessment, publishing, compliance, school administration, recognized qualifications, proprietary data and deeply embedded institutional tools can all possess forms of value that are harder for a general-purpose AI interface to intercept.

That does not make infrastructure businesses invulnerable, and it does not make destination businesses worthless. But the distinction matters because AI dramatically reduces the friction involved in obtaining information, explanation, practice and increasingly personalized instruction without visiting a specialized platform.

For an investor, the question is therefore not merely whether people will continue learning. They obviously will. The question is who captures the economics of that learning when the interface through which it occurs changes.

The public-market comparison makes that problem difficult to ignore. Following the original market hierarchy of the largest ex-China listed education companies turned $1 million into approximately $1.02 million. Over the same period, the same $1 million in SPY became approximately $2.10 million. The specialist investor had to understand pedagogy, procurement, implementation, regulation and a collection of highly specific business models, while the passive investor could effectively buy an index fund and go to lunch. The passive investor ended with more than twice as much money.

That is not an argument that education has no value. It is almost the opposite. People are still learning. Children still need help with anxiety. Adults still need to understand investing. Employees still need new skills. Students still need explanations. The demand is not disappearing. The uncomfortable question is why so many of the companies built to serve that demand have been so poor at turning it into shareholder value.

The MYiNNERME example makes that distinction unusually easy to see. The company can disappear while the knowledge it created, the problems it addressed and the people looking for those solutions remain. That value can now be wrapped in a YouTube universe, recurring characters, AI interactions, books, school programmes, tools for psychologists, physical products or combinations of all of them, often at production and distribution costs that would have been unrealistic only a few years ago.

The movie did not become worthless when the DVD did. But owning a DVD business after streaming arrived was not particularly helpful either.

That is the distinction I think matters for edtech investors now. Good content is not enough. Good pedagogy is not enough. Even enormous demand is not enough. If technology changes how people consume the underlying value, the company has to move with it, and it has to find a way to capture enough of that value to justify the capital invested.

Otherwise, why take the additional risk? If an investor can put $1 million into a passive index, do almost nothing and end with more than twice the money, specialist edtech capital has to earn its existence.

The next generation of winners therefore will not simply be the companies that use AI, the companies with the best courses or the companies that make today's apps marginally better. They will be the ones that understand what part of their product remains genuinely scarce and valuable, how new technology changes the cheapest and most natural way to deliver it, and where in that new system there is actually a business capable of capturing enough value to produce attractive returns.

That is why I think 90% of edtech can disappear without learning disappearing with it. We are not running out of things to learn. We may simply be running out of reasons to pay for them in the ways edtech spent the last decade building.

Series · How 90% of Edtech Disappears

A four-part series on the money, the market, and the models that ended an industry as we knew it.

Part 1 of 4If I Operated an Edtech Fund, I Would Be Shitting My Pants

Part 2 of 4 · Previous in seriesYour Edtech Investor Wants an Open Relationship

Part 3 of 4Edtech Is Still Mailing DVDs

Part 4 of 4 · Next in seriesWhen Your Customer Becomes Your Competitor: Your Customer Found a Shortcut to the Software Factory

Read the whole series →Related essays

AI & Edtech · July 18, 2026If I See One More Person Push “PedTech,” I’m Going to VomitPedagogy was never the missing idea. The harder problem was building an industry whose economics allowed pedagogy to stay at the center once investment dollars came knocking.

AI & Society · September 8, 2026Elon Musk Bought the Wrong Real Estate for the Everything AppMusk bought Twitter to build the everything app. But while he was accumulating attention, OpenAI was accumulating intent, and that may be the most valuable address on the internet.

AI & Society · August 16, 2026AI Ate the Internet. Now We Want It to Decide What’s Human.Watermarking is supposed to make synthetic content transparent. It may also turn AI companies into the institutions we ask to certify human authorship.

← Back to BlogRead part oneSpeakingInvite me to speak

---

### https://sahra-josephine.com/blog/edtech-open-relationship

← Blog

AI & Edtech·July 30, 2026Your Edtech Investor Wants an Open Relationship

How 90% of edtech disappears. You are still in the portfolio. You are simply no longer their only type.

The strangest thing about being slowly dumped is that nobody necessarily ends the relationship. They still answer your messages. Your photograph may still be on the shelf. Their toothbrush may still be in your bathroom. But the language changes. They stop talking about the future in quite the same way, their description of what they are looking for becomes broader, and eventually you realize they have changed their type without formally telling you.

That is what is happening between edtech and its investors.

In my last essay, I predicted that between 90% and 95% of today's edtech companies will close, consolidate or become functionally irrelevant. That prediction rests on several different changes that cannot be explained properly in one essay, so I am going to examine them separately.

I expected to begin with collapsing valuations or artificial intelligence replacing educational products. Instead, I found something quieter and, in some ways, more revealing. Edtech has not necessarily been dumped. The investor simply wants an open relationship.

The money left first

Global edtech venture investment reached $20.8 billion in 2021. By 2024, it had fallen to $2.4 billion, a decline of approximately 89%. It recovered only slightly to $2.6 billion in 2025.

Artificial intelligence did not cause that entire collapse. The decline began before generative AI was widely adopted, and much of it was an inevitable correction after pandemic demand, cheap capital and extraordinary valuations produced an investment boom that could not last.

But four years after the peak, conventional edtech investment remains close to the lowest level recorded in a decade. Capital has not simply returned under the same conditions. It has become more selective, more closely connected to employment and more interested in products that can be described as infrastructure, productivity or artificial intelligence. Then the names began to change.

Edtech changed its dating profile

I reviewed the current public positioning of ten investors historically associated with education and edtech. I selected them because of that association, not because their current language supported my argument. They are not a statistically representative sample of the entire venture market, and a website cannot reveal every decision made by an investment committee. But websites do reveal how funds want founders, portfolio companies and their own investors to understand them.

The pattern was not universal. Owl Ventures, GSV Ventures and Rethink Education remain clearly anchored in education. Educapital still describes itself as an edtech and future-of-work fund. This matters because I was not interested in selecting only the investors that had moved away from the word edtech.

But several of the most recognizable specialist investors now describe a much larger relationship. Reach Capital began in 2015 by investing at the intersection of technology and education. Its renewed thesis covers learning, health and work.

Brighteye no longer leads by calling itself an edtech investor. It backs founders building what it calls the "HumanOS," systems that help people learn, work and adapt continuously. Its flagship market report is no longer called an edtech funding report. It is now the Learning & Work Funding Report.

Emerge was founded as Europe's only specialist edtech fund. It now describes itself as a fund for the future of work and learning, investing in everything from early childhood to career navigation and the use of AI at work.

Kaizenvest says that what began as India's first education-focused private equity fund has evolved into a "comprehensive human economic mobility strategy." Its current thesis stretches across education, healthcare, financial inclusion and job creation.

Learn Capital still speaks extensively about education, but it now describes its territory through education and human capital development, including AI-powered learning, workforce training, career advancement and well-being.

The wording varies, but the direction is remarkably consistent. Edtech becomes learning. Learning becomes skills. Skills become workforce development. Workforce development becomes human capital, career mobility, productivity or human potential. Education is still present. It is simply no longer alone.

This is strategically rational. A dedicated edtech fund must find attractive investments inside edtech. A broader investment thesis can follow the same learner into employment, healthcare, recruitment, productivity or AI. It can preserve its expertise while dramatically expanding the number of companies and budgets it can pursue.

From the investor's perspective, it is diversification. From the founder's perspective, it means that the next cheque no longer has to go to someone like you.

Then I checked my own investor

Only after seeing the broader pattern did I look properly at Sparkmind, one of CanopyLAB's investors. I am not observing this industry as an outsider. I have spent more than ten years building CanopyLAB, which sits in one of the parts of edtech most exposed to AI substitution: platforms that create, organise and deliver learning content.

In 2020, Sparkmind was described as a Nordic venture capital fund specialising in edtech. It planned to invest across the educational journey, from early childhood to lifelong learning and corporate training.

Visit Sparkmind's website today and you are presented with two doors: Human Capital and Security. CanopyLAB and the rest of the original education portfolio remain under Human Capital. That fund has invested in 26 companies and is no longer making new initial investments, which may simply reflect the ordinary life cycle of a venture fund. Sparkmind says it will continue supporting its portfolio companies.

From Sparkmind's perspective, the logic is clear. Education, employment, AI and geopolitical security are all being transformed simultaneously, and a broader mandate creates more places to invest. But strategic logic does not make the founder-side consequence neutral. As a founder, the sensation is unmistakable. I am still in the portfolio, yet edtech is no longer the word that explains the firm's future. The relationship continues, but the investor has changed its type.

The recovery appears when the category gets wider

The new language would be less interesting if it were only branding. The funding data suggests something more substantial. In Europe, investment in the broader "Learning & Work" category more than doubled from €710 million in 2024 to €1.6 billion in 2025. That sounds like a dramatic recovery.

But Brighteye's own breakdown shows that conventional edtech, covering schools, higher education and individual lifelong learning, received €471 million. Corporate and workplace learning received €601 million, while the expanded category also includes productivity, recruitment and talent platforms that do not primarily describe themselves as learning products.

This does not make the larger number wrong. Learning genuinely is becoming integrated with work and performance. But it changes what the recovery means.

Edtech funding remains severely depressed. Learning & Work funding is recovering. The recovery becomes much more impressive after the category expands to include companies that edtech investors would not necessarily have considered edtech five years ago. The market may not be returning to edtech. Edtech may be moving to wherever the market has gone.

Then edtech changes itself to stay desirable

Public companies are changing their language too. Guild Education became Guild, placing career mobility at the center of its story. Chegg now calls its academic business "legacy Academic Services", presents Skilling as its growth engine and has expanded into AI model training. Coursera described its combination with Udemy as a comprehensive skills platform for the AI era, while Multiverse now calls itself Europe's AI adoption platform. The products may still teach people, but the pitch increasingly sells career mobility, productivity, workforce performance or AI adoption. These are not identical decisions, and they do not prove that any particular company will succeed. Collectively, however, they show education becoming less prominent in the language companies use to attract customers and capital.

Public companies make category migration relatively easy to observe. Their investor presentations, earnings calls, acquisitions and changing business segments are visible. Private companies are more difficult. Their valuations are not continuously tested by the market, and the strategic conversations between founders, boards and investors remain private. Their websites are therefore especially revealing. They show which part of the company has been chosen for the shop window.

After looking at Sparkmind, I became curious about the companies beside CanopyLAB in the portfolio. Two of them illustrate very different forms of evolution. Neither case proves that the company will survive. Both show how private edtech companies are already changing their public identities and their positions in the technology stack.

Female Invest no longer chooses the edtech category

Sparkmind classifies Female Invest as a lifelong-learning company. In 2021, it was described as an "EdTech platform and community" using subscription-based learning to help women understand personal finance and investing.

The product is still unmistakably educational. Members receive courses, short lessons, financial news, budgeting tools, access to experts and a virtual trading simulator. What has changed is the category Female Invest chooses to present to the market.

Visit Female Invest's website today and the headline is not about edtech, digital learning or course completion. Female Invest calls itself "the money app for every step of your journey." It tells women to take control of their money, practice investing and build wealth. The community, which the company says includes more than 85,000 women from 125 countries, is presented as part of a movement to close the financial gender gap.

The distinction matters. Female Invest is not simply selling a better way to learn about investing. It is selling confidence, financial independence, expert access, identity and participation in a community built around an unresolved social problem.

Its category has been expanding for years. In 2022, Female Invest acquired Gaia Investments, a sustainable investment platform, with plans to integrate actual trading into its product. The current public offering appears to focus on virtual trading and education rather than operating as a live brokerage, so I would not describe it today as an investment platform. But the acquisition revealed the strategic ambition to stretch from teaching people about money towards helping them act on what they learned.

In 2023, Female Invest explained that its new brand identity would expand beyond financial issues and become a stronger voice on gender inequality more broadly. In 2024, it announced an $11 million Series A. Some coverage called it an edtech company. Other coverage called it fintech. Its current branding makes both labels feel incomplete.

That appears deliberate. Female Invest does not need customers to decide whether it belongs in fintech, edtech, media or community. The brand is organized around the problem it wants to solve rather than the software category used to solve it.

Female Invest consists of practical tools that bring the user closer to action. The simulator allows someone to practice before taking a real financial risk. The community provides social reinforcement. The experts create trust in an area where mistakes have consequences. Together, those elements may create a relationship that is harder to substitute than a conventional catalogue of financial courses.

But none of that proves Female Invest will survive. Educational content, market news and introductory guidance are all highly exposed to AI substitution. A community is only defensible if people participate in it, value the relationships and cannot easily recreate the same experience elsewhere. Female Invest is therefore not evidence that purpose and community guarantee survival. It is evidence that a private edtech company can make education only one part of a much larger identity. The company may continue growing while the category used by its investor to describe it becomes increasingly irrelevant to the people buying the product.

imagi is becoming the layer between schools and AI

The second case is imagi, formerly known more visibly as imagiLabs. In 2022, imagiLabs described itself as an edtech startup launching a platform for instructors to teach Python. Its ecosystem included a gamified mobile app, a wearable imagiCharm and a curriculum designed around the interests of pre-teen girls. The proposition was relatively easy to understand: imagi had built its own environment to help children learn how to code. That is no longer the center of the story.

In July 2026, imagi announced a $4.5 million funding round to build what it calls "the safe education layer between AI tools and schools". Instead of asking students to learn exclusively inside an imagi product, the company now gives them supervised access to general AI tools through curriculum, teacher support and safety controls.

The partners are Lovable and OpenAI. Through the Lovable and imagi collaboration, students use AI to build applications while imagi provides classroom access, lesson plans, automatic student setup and teacher training. The company says it intends to integrate additional frontier AI tools.

This is an almost literal example of the change I described in my previous essay. The learning experience is moving into the general AI environment. The edtech company is repositioning itself as the pedagogical, administrative and safety layer between that environment and the learner.

The change is not merely cosmetic. imagi has moved from trying to own the complete destination to managing access to technologies owned by much larger companies. It is no longer only teaching Python through its own interface. It is helping schools decide how children can safely learn with tools that students may already be using elsewhere.

The early numbers are promising, although they are company-reported. imagi says it operates across more than 100 school districts, has reached more than 700,000 students in 140 countries, increased users thirtyfold over the previous year and tripled annual recurring revenue. Its new funding round included Sparkmind alongside Morgan Stanley and individual investors from companies including ElevenLabs, GitHub, Spotify and Lovable.

Again, this does not prove that imagi will survive. It does, however, reveal why the new position may be attractive.

Schools need more than access to a model. They need child safety, privacy, compliance, classroom management, teacher training, curriculum and someone prepared to take responsibility when something goes wrong. Frontier AI companies may not want to customize those elements for every school district, age group and regulatory environment. imagi is attempting to occupy that gap.

If it can own trusted school relationships, pedagogy, implementation and the ability to connect several AI providers, it may become valuable infrastructure. The company would not need to build the most powerful model. It would need to become the safest and most useful way for schools to access whichever models become important.

But the dependency is obvious. OpenAI, Google, Anthropic, Microsoft, Lovable or one of the large school-platform providers could build more of this layer themselves. They have greater distribution, more capital and control over the underlying technology. I also doubt that the market needs hundreds of independent companies sitting between schools and frontier AI. A few may build defensible positions around geography, age groups, regulation or specific pedagogical needs. Many others may become replaceable wrappers around products they do not control.

That is what makes imagi such an interesting case. Its evolution may be strategically intelligent and entirely consistent with its original mission of helping children, particularly girls, become creators of technology. But the architecture of the company has changed. It now depends on the general AI ecosystem it once might have expected to compete with.

The company has not abandoned education. It has accepted that education may increasingly happen somewhere else, and is trying to become the layer that makes that possible.

Female Invest and imagi do not tell us which companies will survive. They show us how private companies are already responding to the same forces visible among public companies and investors. One has made the category subordinate to purpose, identity and community. The other has moved from being the learning destination to becoming infrastructure around somebody else's AI platform. Both may prove to be excellent decisions. Both may fail. The evidence is not the outcome. The evidence is the movement.

An industry can disappear without everyone dying

I originally imagined the disappearance of edtech as a succession of visible failures: companies closing, valuations collapsing and platforms becoming obsolete. All of that will happen. But industries can also disappear through absorption.

The company survives, but becomes a workforce company. The product survives, but learning becomes one feature inside an AI or productivity platform. The investor survives, but education becomes one possible expression of a much broader investment thesis.

Venture capital was never a marriage. It was never designed for better or for worse. The investor was always going to follow the strongest returns and eventually find an exit. The founder may have entered for a very different reason.

Most people do not spend years building in education because they want to maximize a category multiple. They begin because they care about students, teachers, pedagogy, access or human potential. They want to inspire people, expand opportunity and help someone realize what they are capable of becoming. But once enough external capital is invested, the company acquires another purpose: protecting shareholder value.

When edtech stops attracting capital, that pressure can send a company away from its core. Education becomes skills. Learning becomes productivity. Students become human capital. The original mission remains on the About page while the commercial center moves towards whoever still has money.

Sometimes that is legitimate adaptation. Sometimes it is the only way to survive. But product-market fit should tell a company how to deliver its purpose. It should not be allowed to decide what that purpose is.

Venture capital moves on. The edtech company remains, changing itself into whoever the market might love next. That is how edtech disappears: not when every company dies, but when staying alive requires forgetting why it was born.

Series · How 90% of Edtech Disappears

A four-part series on the money, the market, and the models that ended an industry as we knew it.

Part 1 of 4 · Previous in seriesIf I Operated an Edtech Fund, I Would Be Shitting My Pants

Part 2 of 4Your Edtech Investor Wants an Open Relationship

Part 3 of 4 · Next in seriesEdtech Is Still Mailing DVDs

Part 4 of 4When Your Customer Becomes Your Competitor: Your Customer Found a Shortcut to the Software Factory

Read the whole series →Related essays

AI & Edtech · July 18, 2026If I See One More Person Push “PedTech,” I’m Going to VomitPedagogy was never the missing idea. The harder problem was building an industry whose economics allowed pedagogy to stay at the center once investment dollars came knocking.

AI & Society · September 8, 2026Elon Musk Bought the Wrong Real Estate for the Everything AppMusk bought Twitter to build the everything app. But while he was accumulating attention, OpenAI was accumulating intent, and that may be the most valuable address on the internet.

AI & Society · August 16, 2026AI Ate the Internet. Now We Want It to Decide What’s Human.Watermarking is supposed to make synthetic content transparent. It may also turn AI companies into the institutions we ask to certify human authorship.

← Back to BlogRead the first essaySpeakingInvite me to speak

---

### https://sahra-josephine.com/blog/edtech-fund

← Blog

AI & Edtech·July 25, 2026If I Operated an Edtech Fund, I Would Be Shitting My Pants

I have spent more than ten years building in edtech.

I co-founded CanopyLAB, a venture-backed learning technology company, because I believed adaptive learning would fundamentally change education. I still believe I was right about that.

What I underestimated was how long it would take for the technology to catch up with the educational principles.

For years, adaptive learning was more convincing as an idea than as an actual user experience. We could collect data, recommend content, and design different pathways through a course. But creating and maintaining those pathways required too much manual work from instructors. The technology could support adaptation, but it could not truly understand the learner, generate the right material, and reshape the experience continuously.

It was adaptive, but within boundaries someone else had already created.

That started to change with artificial intelligence.

In 2019, I invented AICATO, which CanopyLAB launched as the world's first AI course authoring tool. It could automate parts of course creation that previously required substantial human effort. It was an early attempt to solve one of the fundamental problems in adaptive learning: you cannot create a genuinely individual learning experience if every piece of content and every possible pathway must first be produced by hand.

Then generative AI arrived, and suddenly the technology could do much more than select between predefined options. It could generate, explain, question, adjust, and respond in real time. The technology finally caught up with the idea.

That should make me extraordinarily optimistic about edtech. Instead, it makes me nervous.

I am giving companies the opposite advice today

Ten years ago, if a company asked whether it should build its own learning platform, I would almost always have said no. Building software was expensive. It took time. You needed specialist developers, designers, product managers, infrastructure, and continuous maintenance. Even if you managed to build something functional, it would rarely be as good as a product developed by a company that did nothing else.

The sensible advice was to find a specialist provider. Run a tender. Buy the best available product. Customize it where necessary. Today, I increasingly give companies the opposite advice. This is not limited to edtech. I have just under 100,000 subscribers across my English- and Spanish-speaking email lists. Maintaining and sending to the 36,000 people on my Spanish language list through Mailchimp alone was costing me around 10,000 Danish kroner every month (1,520 USD). That is a significant amount of money to send email to a list I already own. So last week, I used Lovable to build my own sending tool and connected it to SendGrid. Building the tool cost me less than $30. Operating it costs approximately $25 to $30 a month. Why would I continue paying for the software subscription?

Of course, not every company should build every tool. Small companies may still benefit from buying something ready to use. Complex and regulated systems are a different matter. But once you have 30,000, 40,000 or 100,000 users, the calculation begins to change dramatically.

A tool that once required a software company, a development team and venture capital can increasingly be created by one person over a weekend. In some cases, a company can recover its development investment after one or two months of cancelled subscription fees. This challenges one of the central assumptions behind the software as a service economy. We spent years moving from ownership to rental because specialist software was cheaper and better than anything most organizations could build themselves. Now the cost of building is collapsing, while the cost of renting has not.

Edtech has a second and even bigger problem

Edtech is not only vulnerable because organizations can build more of their own software. It is also vulnerable because learners may no longer need a separate learning platform at all. Elon Musk has spoken for years about turning X into an everything app. I think the fundamental idea was right, but the chosen starting point was wrong.

The everything app was probably never going to be a social network. It is going to be the AI platform that already knows what you are working on, what you are interested in, what you understand, and where you repeatedly get stuck. That platform could be ChatGPT, Claude, Gemini, Grok, or something that has not yet been released. The winner is not necessarily important to my argument. But the shift is already happening.

When I want to understand how to maximize the return from my investment portfolio, I do not necessarily want to enrol in an online course about investing. I can begin a conversation with an AI platform immediately. It already knows a great deal about my level of knowledge, my goals, the questions I have asked and the decisions I am trying to make. It can explain a concept, test whether I understand it, adjust the difficulty, generate examples based on my actual portfolio, and move in a completely different direction when my needs change. That is adaptive learning. It is more adaptive than most products that describe themselves as adaptive learning platforms because it is not adapting my journey through a fixed course. It is creating the learning experience around me as I go.

An edtech platform ordinarily begins with no knowledge of me. It asks me to create another account, complete another onboarding process, and perhaps take a diagnostic test. Even then, it only knows what I have done inside that particular platform. My preferred AI platform may already know what I have been reading, writing, building, and struggling with across my life and work. How is a conventional learning platform supposed to compete with that?

Chegg is not the story. It is the warning

Chegg is one of the clearest public examples of what happens when learners stop going to a dedicated education product for answers. At its pandemic peak, Chegg was valued at close to $15 billion. By July 2026, its value had fallen to around $100 million. The demand for help did not disappear. Students still needed explanations and assistance. What disappeared was the reason to pay Chegg for access to them.

Generative AI can now answer many of the same questions people turned to Chegg for, instantly. Chegg itself acknowledged that generative AI and Google's AI Overviews were reducing traffic and subscriptions. The company subsequently made enormous reductions to its workforce. But Chegg is publicly traded. We can see the collapse in its share price.

If I operated an edtech fund, I would be more concerned about the companies whose decline is not visible yet. Private companies remain recorded at valuations established during previous financing rounds. Institutional contracts can take years to expire. Universities and corporations move slowly. Revenue can therefore continue while the underlying reason for the product to exist is disappearing.

A portfolio can look healthy on paper long after its investment thesis has stopped being true. Adding an AI assistant to the existing platform does not solve that problem. The question is not whether an edtech company uses AI. Nearly all of them will. The question is why a learner would enter that platform instead of simply learning inside the AI environment they already use for everything else.

What can survive?

I do not believe all edtech will disappear. But I believe that between 90% and 95% of today's edtech companies will eventually close, consolidate, or become functionally irrelevant. And this will happen within the next 2–3 years.

The survivors will need to offer something that cannot be reproduced by asking an AI platform to teach you. That could be a real community. A strong sense of identity and belonging. Access to people you genuinely want to learn with or from. A credential that has meaningful value. A physical or social experience. Or a strong ideological foundation that makes participation mean something beyond acquiring information.

People do not join Harvard only because Harvard possesses information unavailable elsewhere. They join because of what Harvard represents, who else is there, and what membership unlocks. The same distinction will increasingly determine which learning companies survive. Information is no longer enough. Content is no longer enough. Personalization is no longer enough. Even adaptive learning is no longer enough because the general AI platforms may be able to do it better, with more context and without asking the learner to start over.

The technology finally arrived. It may destroy the category built to deliver it

For more than a decade, I believed adaptive learning represented the future of education. I was right. But I had assumed that adaptive learning would transform learning platforms. I now think it may replace many of them. The biggest threat to edtech is not another edtech company with a better feature set. It is that learning becomes one behaviour among many inside a much larger AI environment.

At the same time, the organizations buying learning technology are discovering that they can build more of it themselves. That leaves traditional edtech squeezed from both directions. Its customers may no longer need to buy the software. Its learners may no longer need to visit the platform. So when shareholders, employees and other people ask me where I think the space is going, I cannot give them the reassuring answer they probably expect from someone who has spent more than ten years building in it.

I think education is entering one of the most exciting periods in its history. But if my investment portfolio consisted of companies built around delivering it through separate software platforms, I would be shitting my pants.

Series · How 90% of Edtech Disappears

A four-part series on the money, the market, and the models that ended an industry as we knew it.

Part 1 of 4If I Operated an Edtech Fund, I Would Be Shitting My Pants

Part 2 of 4 · Next in seriesYour Edtech Investor Wants an Open Relationship

Part 3 of 4Edtech Is Still Mailing DVDs

Part 4 of 4When Your Customer Becomes Your Competitor: Your Customer Found a Shortcut to the Software Factory

Read the whole series →Related essays

AI & Edtech · July 18, 2026If I See One More Person Push “PedTech,” I’m Going to VomitPedagogy was never the missing idea. The harder problem was building an industry whose economics allowed pedagogy to stay at the center once investment dollars came knocking.

AI & Society · September 8, 2026Elon Musk Bought the Wrong Real Estate for the Everything AppMusk bought Twitter to build the everything app. But while he was accumulating attention, OpenAI was accumulating intent, and that may be the most valuable address on the internet.

AI & Society · August 16, 2026AI Ate the Internet. Now We Want It to Decide What’s Human.Watermarking is supposed to make synthetic content transparent. It may also turn AI companies into the institutions we ask to certify human authorship.

← Back to BlogAbout Sahra-JosephineSpeakingInvite me to speak

---

### https://sahra-josephine.com/blog/pedtech

← Blog

AI & Edtech·July 18, 2026If I See One More Person Push “PedTech,” I’m Going to Vomit

Pedagogy was never the missing idea. The harder problem was building an industry whose economics allowed pedagogy to remain at the center when investment dollars came knocking.

I have apparently reached the stage of my career where ideas we discussed as novel more than ten years ago have now returned with new names and are presented to me as revolutionary discoveries. Watching the current conversation around “PedTech,” I keep thinking about The Emperor's New Clothes. Everyone is admiring the new outfit, and I am standing in the corner wondering whether we are allowed to point out that we have seen it before.

The latest outfit is “PedTech.” The argument seems sensible enough. We have focused too much on technology in the education space and need to redesign around pedagogy, ensuring better learning outcomes and designing for actual improved learning. Start with the learning problem. Understand the learner. Support the teacher. Stop digitizing content for the sake of digitizing content. Build technology around how people actually learn.

Yes, obviously. We were saying this ten years ago, and people were saying it long before I entered the industry. Teachers were saying it before most edtech companies existed. Researchers built entire careers around it. Instructional designers had methodologies for it, and edtech founders wrote it into their pitch decks. So when I see “pedagogy first” presented as the correction to a previous generation of edtech, I find the implied history slightly insulting.

I have never hired anyone to work on the educational side of CanopyLAB who wasn't deeply passionate about learning and pedagogy. The problem was not that an entire generation forgot pedagogy. The problem is much more uncomfortable: a great many people who cared deeply about pedagogy built companies inside an economic system that repeatedly made it difficult to keep pedagogy first. That distinction matters.

Nobody chose edtech because it was the easiest place to get rich

There is a strange caricature beginning to emerge of the previous generation of edtech. Apparently, we were a collection of technologists who became so excited about software that nobody remembered to ask whether students learned anything. I do not recognize that industry.

Of course there were opportunists. There were bad products, founders who over time cared more about growth than outcomes, and investors who cared primarily about returns. That is true of every technology category. But education technology has never been the most obvious place to go if your sole objective is maximizing financial return.

When I told my friends I was leaving academia to become a founder in tech, they were excited. When I told them that I was building an edtech company, they reminded me that before academia I had been in consulting, and that returning there probably made more sense financially and reputationally.

If I knew then what I know now, I would seriously reconsider starting anything in the edtech space. Institutional sales cycles are long. Procurement is painful. Budgets are constrained. Teachers are overloaded. Schools are political. Universities move slowly. Implementation is expensive. Education systems are fragmented across countries, municipalities, institutions, curricula, languages and regulatory environments. The person using the product is frequently not the person buying it, and the person buying it may not control whether anybody uses it. Even when the product works beautifully, proving that somebody actually learned because of it can be surprisingly difficult.

There have always been easier places to sell software. A great many people entered education technology because they cared about education. Teachers became founders. Researchers became founders. Parents became founders. People who had experienced terrible education systems became founders. People who believed access could be improved became founders. People who thought technology could give learners something the existing system could not became founders. Many of us were idealistic enough to believe that better pedagogy and better technology could reinforce each other. That did not prevent us from building an industry with serious problems, but it changes the diagnosis.

Caring about pedagogy did not save us

This is the part I think the PedTech conversation risks missing. Good intentions do not automatically produce good systems. A founder can care deeply about teachers and still build software teachers do not use. A company can begin with excellent learning science and still end up optimizing completion rates. An instructional designer can create a beautiful learning experience and discover that the buyer wants 147 features in the procurement document. A team can believe in learner agency and still discover that its largest customer wants mandatory pathways, reporting and administrator controls.

A founder can want to reduce teacher workload and still be forced to add another dashboard because the person signing the contract needs something measurable to show their boss. A company can genuinely care about learning outcomes and still discover that logins, clicks, minutes watched and courses completed are dramatically easier to put into a quarterly report.

Nobody has to be evil for this to happen. The incentives only have to point in slightly different directions, because over enough years slightly different directions become entirely different destinations. That is why I find “we need to put pedagogy first” insufficient. Many people tried. The more useful question is why they so often failed to keep it there.

Perhaps the failure was structural, not philosophical

Imagine that the entire previous generation had been given better advice: pedagogy first, diagnose the problem, understand the learner, support the teacher and measure outcomes. Would the industry necessarily look dramatically different? I am not convinced.

Eventually somebody still has to pay. That person may be a school district, a university, a corporation, a ministry, a parent or an investor, and each introduces constraints that have very little to do with instructional theory. The school district wants integration. The university wants compliance. The corporation wants reporting. The procurement team wants security documentation. The investor wants growth. The board wants recurring revenue. The founder wants enough runway to survive another year.

Then the beautifully diagnosed pedagogical problem enters a commercial machine, and that machine changes things. Perhaps the best intervention requires substantial teacher involvement, which makes implementation expensive. Perhaps the learner should only use the product occasionally, which makes retention look terrible. Perhaps the correct solution is five excellent learning moments rather than fifty hours of content, which makes the catalogue look small. Perhaps the most pedagogically responsible conclusion is that the institution does not need another platform, which makes the sales meeting awkward.

The conflict was never simply pedagogy versus technology. It was pedagogy versus the economics surrounding the technology.

If pedagogy really comes first, sometimes there is no software company

This is the test that interests me. Imagine somebody diagnoses an educational problem perfectly. They understand the learner, the teacher, the context and what the evidence suggests is likely to work. After doing all of that, they discover that the best intervention is a teacher guide, three printable exercises and a weekly conversation. Wonderful. Is that PedTech? Probably not. Is it pedagogically sound? It might be.

Now imagine another problem where the best intervention is a WhatsApp group, a workbook and access to a human mentor. Again, perhaps excellent education, but still not a particularly exciting venture-backed software company. This is where “pedagogy first” becomes much more demanding than it initially sounds. Pedagogy does not owe us software. It does not owe us recurring revenue or venture returns, and it certainly does not care whether the intervention has gross margins above 80%.

If we genuinely put pedagogy first, we have to accept that sometimes technology comes second, sometimes it comes last and occasionally it should not come at all. That creates a difficult problem for an industry whose business model requires technology to remain somewhere near the center.

And the timing of the rebrand is interesting

I might be less irritated by PedTech if it were appearing in a vacuum, but it is not. Edtech is going through a category identity crisis at precisely the same time as people are discovering that perhaps they were never really edtech companies or investors in the first place.

Investors historically associated with education increasingly describe themselves through broader territories. I have argued before that edtech becomes learning, learning becomes skills, skills become workforce development, and workforce development becomes human capital, career mobility, productivity, AI adoption or human potential. Companies are doing the same thing. Education companies become skills companies. Learning companies become workforce platforms. Financial education companies become money apps. Coding products become AI adoption products.

Many of these changes make perfectly good strategic sense. Markets move, technology changes and companies should evolve. I have no particular attachment to keeping a company inside a category that no longer describes what it does. But the timing is worth noticing.

Edtech has endured collapsing venture investment, disappointing public-market outcomes, low utilization of software and now competition from general-purpose AI products capable of delivering many of the things that previously required a specialized platform. At precisely that moment, a remarkable number of people are discovering that they no longer belong to edtech. Then, right in the middle of that identity crisis, we discover PedTech.

Perhaps it represents a genuinely new discipline, and I am open to being convinced. But when a struggling category suddenly discovers a new name for one of its oldest principles, I think we are entitled to inspect the clothes before applauding the emperor.

You cannot rebrand your way out of incentives

Suppose every edtech company changed its name to PedTech tomorrow. What would actually be different on Monday morning? The school procurement department would send the same tender. The corporate buyer would request the same reporting. The teacher would still have the same number of hours in the day. The investor would still expect a return, the board would still track recurring revenue, the sales team would still have a quota and the customer would still expect implementation.

Nothing important would have changed, which is why I am much more interested in whether PedTech represents a different economic and delivery model than whether it represents a different philosophy. If customers begin paying for demonstrated outcomes rather than access, that is interesting. If teachers actually gain time instead of receiving another administrative responsibility, that is interesting. If evidence becomes central to procurement, products become smaller because smaller is pedagogically better, or investors become comfortable funding businesses whose educational value does not naturally translate into conventional SaaS economics, then we are beginning to talk about structural change.

Those changes would matter because they alter the incentives surrounding the pedagogy. Putting “pedagogy first” on a slide does not.

The previous generation deserves criticism. There were too many platforms, too many unused licenses, too much content, too little evidence and too many engagement metrics mistaken for learning. Many products required enormous behavioral change from teachers who were already exhausted, and too much capital went into companies whose economics probably never justified it. Some of us built that industry, and we should be willing to admit where we were wrong.

But if people with good intentions, relevant expertise and a genuine commitment to education still produced products with poor adoption, weak evidence and difficult economics, then the interesting question is not whether they remembered to put pedagogy first. It is why caring about pedagogy was so often insufficient. Maybe procurement rewards features more reliably than outcomes. Maybe venture capital pushes education companies toward growth models that do not fit education. Maybe SaaS encourages recurring usage even when good learning does not require recurring software usage. Maybe institutional buyers and individual learners value fundamentally different things. Maybe some educational problems are simply not good venture-backed software opportunities. Those questions are considerably harder than inventing a new category, which is precisely why they are more useful.

A new category should explain something new

I am not opposed to new terminology. Useful categories help us distinguish things that previously looked the same because something material has changed. SaaS described a genuinely different delivery and economic model, while generative AI describes a genuinely different technological capability. A useful category should therefore be able to tell us what has changed.

So tell me what changed with PedTech. What can a company do today that a pedagogy-led education company could not do ten years ago? What new delivery model exists? What new economic model exists? What incentive has been realigned? What institutional constraint disappeared? What evidence standard changed, and what changed in the behavior of the buyer?

If PedTech answers those questions, I will happily listen. But if the answer is simply that we should start with pedagogy, diagnose the learning problem, understand the learner and support teachers, then I am sorry: you have not invented a category. You have described what good edtech was always trying to become.

That distinction matters because if the previous generation failed primarily because it had the wrong philosophy, the solution is wonderfully easy. Teach everyone the correct philosophy, give it a new name and try again. If the failure was structural, the work ahead is much harder. We have to rethink what gets funded, what gets purchased, what gets measured, how learning is delivered, where software belongs and whether every valuable educational intervention needs to become a scalable technology company in the first place.

That would be a real reset. PedTech, so far, is a word. We should absolutely put pedagogy first, and we should have done it yesterday as well as tomorrow, but we should not confuse rediscovering an old principle with changing the system that repeatedly made the principle difficult to follow.

The conversation worth having is not EdTech versus PedTech, which word gets the next conference track or who has suddenly discovered that teachers matter. The question is whether we can finally build an economic and delivery model in which the thing everyone claims to put first is actually allowed to stay there.

If PedTech does that, dress it however you like, and I will happily admit that the emperor has a new wardrobe. But if all we have done is rediscover pedagogy and give it a conference track, somebody eventually has to point out that we have seen these clothes before. If they are presented to me as a revolutionary new outfit one more time, I may genuinely vomit.

Related essays

AI & Edtech · September 5, 2026When Your Customer Becomes Your Competitor: Your Customer Found a Shortcut to the Software FactoryPart four of how 90% of edtech disappears. Software companies spent twenty years convincing customers not to build. Artificial intelligence is changing the distance between wanting software and making it.

AI & Edtech · July 25, 2026If I Operated an Edtech Fund, I Would Be Shitting My PantsAfter a decade building in edtech, I think 90–95% of today's edtech companies will close, consolidate, or become irrelevant within 2–3 years.

AI & Edtech · August 13, 2026Edtech Is Still Mailing DVDsHow 90% of edtech disappears: people have not stopped learning, they have started consuming knowledge differently, and that distinction may wipe out much of the industry.

← Blog

---

## Dansk (automatisk oversat fra engelsk)

### https://sahra-josephine.com/da/blog/everything-app

← Blog

AI & samfund·8. september 2026Elon Musk købte den forkerte ejendom til alt-appen

Denne tekst er maskinoversat fra engelsk. Nuancer kan gå tabt undervejs, så vi anbefaler at læse originalen. Læs den engelske original

I årevis annoncerede han, at X skulle blive appen til alt. OpenAI har måske stille og roligt bygget noget, der kommer tættere på, for den mest værdifulde ejendom er ikke nødvendigvis der, hvor vi bruger vores opmærksomhed. Det er der, hvor vi udtrykker vores intention.

Silicon Valley har brugt de sidste tyve år på at forsøge at placere dig inden i The Truman Show. Ikke i den forstand at Mark Zuckerberg i al hemmelighed vil caste dine naboer og filme dig, mens du sover. I den mere jordnære og sandsynligvis mere indbringende forstand, at de største teknologivirksomheder gang på gang har forsøgt at opbygge en verden, hvor næsten alt, hvad du gør, foregår på grunde, de ejer.

Du taler der. Du handler der. Du ser fjernsyn der. Du bestiller ferien der. Du bestiller aftensmad der. Du sender penge der. Ideelt set har du aldrig en grund til at forlade det. Det er super-app-drømmen.

Elon Musk har været usædvanlig åben omkring det. Han købte Twitter for 44 milliarder dollars, omdøbte det til X og gik i gang med at tilføje betalinger, video, skabere, finansielle tjenester og AI. Han har beskrevet ambitionen helt direkte: alt-appen.

Sidst i juli blev X Money lanceret for udvalgte brugere i USA. Man kan sende penge til andre brugere, få et X-brandet Visa-kort, få renter af sine indskud og begynde at lave den slags ting, der ikke har det fjerneste at gøre med at skændes med fremmede på Twitter. Det er det seneste, meget bogstavelige skridt i Musks årelange forsøg på at opføre sin egen udgave af Seahaven.

Jeg tror, han kan have ret i alt-appen. Jeg tror bare, han købte den forkerte ejendom. For mens Musk højlydt har annonceret, at han agter at bygge en, har OpenAI gjort noget langt mindre teatralsk. Det startede med en chatboks. Så lærte chatboksen at researche. Den lærte at arbejde med filer. Den fik apps. Den lærte at shoppe. Nu kan den navigere på hjemmesider og udføre handlinger på brugerens vegne.

Der var ingen stor meddelelse om, at ChatGPT agtede at blive stedet, hvor man gør alt. Den blev bare ved med at opsuge ting, vi før åbnede andre produkter for at gøre. På et tidspunkt holder det op med at ligne en chatbot med feature creep. Det begynder at ligne alt-appen på uhyggelig vis. Eller i det mindste en langt stærkere udgave af en. Måske behøver virksomheden, der vinder kapløbet om at bygge alt-appen, slet ikke eje hele Trumans verden. Den skal bare blive den person, han spørger om alt.

Musk købte opmærksomhed

Logikken bag at købe Twitter gav mening. Hvis man vil bygge en alt-app, kan man starte et sted, hvor hundredvis af millioner mennesker allerede tilbringer deres tid. Musk købte ejendommen og begyndte at føje flere ting til den. Betalinger. Video. Skabere. AI. Finansielle tjenester. Strategien er i bund og grund at blive ved med at bygge, indtil der er færre grunde til at gå.

Sådan har vi historisk tænkt på værdifuld internetejendom. Amazon ejer stedet, jeg går hen, når jeg vil købe noget. LinkedIn ejer stedet, jeg går hen, når det handler om arbejde. Booking ejer en destination for rejser. YouTube ejer video. Twitter ejede den offentlige samtale. Brugeren beslutter, hvor de vil hen, og fortæller derefter produktet, hvad de vil have. Men AI vender den rækkefølge om. I stigende grad behøver jeg ikke vide, hvor jeg skal starte. Jeg kan begynde med det, jeg forsøger at opnå.

Jeg skal have et sted at bo i Barcelona, men jeg skal af sted til lufthavnen klokken tre. Jeg skal have styr på, om den her kontrakt er elendig. Mit barn forstår ikke brøker. Jeg vil have et par turkise Aquazzura-stiletter, men ikke til 900 dollars. Jeg skal vide, om den her virksomhed faktisk vokser, eller bare udsender entusiastiske pressemeddelelser. Jeg kan begynde med problemet. Destinationen kommer bagefter.

Det lyder som en lille ændring i brugerfladen. Jeg tror, det er en enorm ændring i, hvor magten sidder på internettet. X vil have mig til at gøre alt der. ChatGPT behøver kun, at jeg begynder alt der.

Den bedste adresse på internettet er måske intention

I størstedelen af internettets historie har virksomheder brugt enorme summer på at regne ud, hvad vi vil have. Annoncering er i høj grad en udførlig øvelse i at gætte. Jeg så det her. Jeg klikkede på det der. Jeg søgte på Mexico. Jeg dvælede ved en sofa. Et sted konkluderer adskillige computere, at jeg muligvis er ved at flytte, og bruger den næste måned på at vise mig spiseborde. Google kom meget tættere på den værdifulde del, fordi søgning fanger eksplicit intention. Jeg skriver »hotel Mexico City«, og Google behøver ikke gætte ret meget.

Den forskel gav dramatisk forskellig økonomi. Twitter havde enorm kulturel betydning, men genererede cirka 4,5 milliarder dollars i annonceindtægter i 2021 og sagde, at det udgjorde under tre procent af det digitale annoncemarked. Google Search og relateret annoncering genererede 63,3 milliarder dollars alene i andet kvartal af 2026. Det er ikke en ren sammenligning. Virksomhederne adskiller sig i skala, geografi, produkter og annonceinfrastruktur. Men det illustrerer, hvorfor det øjeblik, hvor nogen udtrykker et behov, har været så værdifuldt et territorium.

Traditionel søgning gav stadig i høj grad beslutningen tilbage til mig. Her er linkene. Her er hotellerne. Her er annoncerne. Held og lykke. Den agentiske version går videre. Jeg fortæller systemet, hvilket resultat jeg vil have, og det kan i stigende grad researche mulighederne, sammenligne dem, anbefale, hvad jeg skal gøre, og udføre det næste skridt.

Google tjente penge på: Hvad vil du finde? En agent kan potentielt tjene penge på: Hvad vil du have gjort? Det er det stykke ejendom, jeg tror, Musk undervurderede: øjeblikket før brugeren vælger virksomheden.

Jeg bringer selv engagementet med

Men den økonomiske sammenligning fanger ikke den mærkeligste forskel mellem X og en LLM. Traditionelle forbrugerplatforme har brug for, at andre hele tiden producerer noget, der er værd at engagere sig i. Facebook har brug for venner. X har brug for folk, der siger noget. YouTube har brug for skabere. TikTok har brug for en ubarmhjertig strøm af videoer, der kan holde på vores opmærksomhed. Uden mennesker, opslag og videoer står ejendommen tom. En LLM kan skabe et langt engagement-loop ud af noget, jeg selv bringer med.

Jeg kan dukke op med et forretningsproblem, en ufærdig idé, en beslutning, jeg ikke kan træffe, et forhold, jeg ikke forstår, eller noget, jeg skammer mig over at spørge et andet menneske om. Der behøver ikke ligge en ny video og vente på mig. Ingen andre behøver at være online. Et andet menneske behøver ikke skabe genstanden for min opmærksomhed, før sessionen kan begynde. Jeg bringer selv råmaterialet til engagementet med.

Det kan allerede ses i måden, folk bruger produktet på. I OpenAIs analyse af forbrugersamtaler i ChatGPT blev 49 procent af beskederne klassificeret som »Asking« — at søge information, vejledning eller råd — mens yderligere 40 procent handlede om »Doing«. Folk kom ikke primært for at forbruge noget, et andet menneske havde lavet. De kom med spørgsmål, beslutninger og ufærdigt arbejde fra deres eget liv.

Den ene model afhænger primært af engagement med andre mennesker og det, de producerer. Den anden kan begynde med et engagement mellem mig, mit eget liv og modellen. Det ændrer, hvad platformen kan vide. Opmærksomhed fortæller en platform, hvad der holdt på mig. Intention fortæller den, hvad jeg vil have. Samtale kan fortælle den hvorfor.

Meta har historisk været nødt til at observere, at jeg så seks videoer om Milano, og udlede, at jeg måske gerne vil dertil. I en samtale kan jeg helt enkelt forklare, at jeg har afsluttet et mangeårigt venskab, gerne vil være et sted med liv, hader store kædehoteller, har brugt for mange penge for nylig og ikke vil gentage min sidste tur dertil. Det er ikke adfærdsaffald fra engagementet. Det er engagementet.

Alt-appen indeholder måske slet ikke alt

Den traditionelle super-app-tese handler om akkumulering. Læg beskeder, betalinger, shopping, underholdning og tjenester ind i ét produkt, indtil det indeholder en så stor del af nogens liv, at det bliver upraktisk at forlade det.

AI-versionen kræver langt mindre ejerskab. Den kan lade flyselskabet forblive et flyselskab og hotellet et hotel. Shopify kan stadig drive butikken. Banken kan stadig opbevare pengene. Uddannelsesvirksomheden kan stadig udstede beviset. De underliggende tjenester behøver ikke forsvinde. De bliver bare ting, agenten kan vælge.

Det er ikke længere kun teori. Produktdata fra millioner af forhandlere er allerede integreret i ChatGPT gennem Shopify Catalog. Forhandleren står stadig for butikken, brandet, betalingen og leveringen. ChatGPT kan indtage samtalen, hvor kunden beskriver behovet og vurderer mulighederne.

Alt-appen er måske ikke stedet, hvor jeg gør alt. Det er måske stedet, hvor jeg forklarer, hvad jeg vil have gjort.

Forbrugeradfærden er ved at følge efter. Adobe fandt, at trafikken fra generative AI-tjenester til amerikanske detailhjemmesider steg 393 procent år til år i første kvartal af 2026. I marts konverterede de besøgende 42 procent bedre end anden trafik. Forhandleren gennemførte stadig handlen. Men en voksende del af beslutningen var allerede truffet et andet sted. Butikken bliver. Besøget bliver valgfrit.

Derfor ser alt-app-idéen anderledes ud gennem en AI-linse. Musks version handler om at bringe mere af internettet ind i X. Den agentiske version kan lade internettet blive, hvor det er, og blive stedet, hvorfra jeg tilgår det. Browseren spurgte, hvor jeg ville hen. Søgningen spurgte, hvad jeg ledte efter. Agenten spørger, hvad jeg forsøger at opnå. Det underliggende internet kan forblive enormt, mens det bliver stadig mere usynligt for den, der bruger det.

Opmærksomhed er ikke forældet

Der er et indlysende problem med at erklære intention for den bedre ejendom. Opmærksomhed kan skabe intention. Jeg åbner ikke Instagram, YouTube eller X, fordi jeg allerede har besluttet, at jeg har brug for et bestemt par sko, en restaurant eller en ferie. Nogle gange vil jeg have tingen, fordi nogen satte den foran mig.

Meta er det stærkest tænkelige argument imod at påstå, at opmærksomhed er økonomisk svag. Det genererede 196,2 milliarder dollars i annonceindtægter i 2025. Men Meta byggede den maskine ved at blive usædvanlig dygtig til at udvinde udledt kommerciel intention af opmærksomhed.

Jo tættere det kommer på at vide, hvad jeg måske køber, jo mere værdifuld bliver opmærksomheden. En LLM kan springe meget af den udledning over. Engagementet indeholder selve intentionen.

Forskellen er derfor ikke, at opmærksomhed betød noget i går, og intention vil betyde noget i morgen. Opmærksomhed er stadig ekstraordinært værdifuld til at skabe efterspørgsel. Intention er ekstraordinært værdifuld til at fordele den. En agent, der kan handle på anmodningen, bevæger sig potentielt endnu et skridt videre, fra eksplicit intention til delegeret intention. Feeds kan være med til at forme lyst. Agenter kan beslutte, hvor den ender.

Der er også en indlysende komplikation: Musk ejer også en AI-virksomhed. Grok er dybt integreret i X, og der findes ingen teknologisk lov, der forhindrer X i at blive langt mere nyttig, når en bruger ankommer med en eksplicit intention. Det her er ikke en nekrolog over X. Det er et spørgsmål om vane. Kan en destination, der primært er bygget omkring opmærksomhed, blive det sted, folk instinktivt begynder, når de vil have noget gjort? Eller har det sted, de allerede beder om at få ting gjort, en lettere vej ind i mere af det, de laver?

Og så er der pengene

X Money er måske det stærkeste argument for Musks version af alt-appen. Betalinger er usædvanlig stærk infrastruktur. Penge rører næsten alt. Hvis X kan blive et sted, hvor folk opbevarer penge, betaler hinanden og bruger et kort, skaber det både ekstraordinær transaktionshyppighed og en relation, der er betydeligt sværere at forlade.

Det er ikke en ubetydelig fordel. Men den afslører også forskellen mellem de to modeller. Musks version siger: Ej også betalingssystemet. Den agentiske version behøver ikke nødvendigvis det. Hvis jeg beder en agent om at købe noget, kan forhandleren forblive forhandler, banken forblive bank og betalingsudbyderen forblive betalingsudbyder. Agenten skal blot have tilladelse til at forbinde dem. Hvilket skaber et mere interessant spørgsmål end om betalinger betyder noget. Er den mere værdifulde position den, der ejer tegnebogen? Eller den, der modtager anmodningen og beslutter, hvor tegnebogen skal sende pengene hen? Det er ikke det samme væddemål.

Agenten får Milano og mig

Der er endnu en grund til, at det betyder noget at blive spurgt først. Intention er, hvad jeg fortæller et system nu. Kontekst er, hvad det allerede ved.

Forestil dig, at jeg beder det finde mig et sted at bo i Milano. Et hotelwebsite kender måske mine datoer, min loyalitetsstatus, tidligere bookinger og de præferencer, jeg eksplicit har indtastet.

En assistent, jeg har brugt i årevis, kan potentielt vide betydeligt mere. Den ved måske, at jeg følger Formel 1 og elsker ballet. At jeg bruger en urimelig stor del af min tid på at tænke over restauranter. At jeg interesserer mig for mode, som regel bor på femstjernede hoteller, er medlem af Soho House og tidligere har afvist anbefalinger, fordi jeg syntes, ejendommene så sjælløse ud. Ingen af de oplysninger er særlig værdifulde hver for sig. Sammen ændrer de betydningen af anmodningen.

Et bookingsite får Milano og mine datoer. Agenten får Milano og mig. Men det skaber den sværeste indvending mod hele argumentet. Brugerfladen, der forstår mine intentioner, kan også påvirke dem. En søgemaskine møder normalt anmodningen, efter at jeg har reduceret den til en forespørgsel. En assistent kan være til stede, mens jeg stadig er i færd med at beslutte, hvad anmodningen er.

Hvis det samme system hjælper mig med at formulere et ønske, afgør hvilke muligheder der fortjener overvejelse, anbefaler en af dem og kan gennemføre handlen, bliver rådgivning og kommerciel indflydelse svære at skille ad. Assistenten kan påstå at forstå, hvad jeg vil have, mens dens forretningsmodel belønner den for at ændre, hvor det ønske ender — eller ligefrem være med til at skabe det.

At eje kontekstualiseret intention er derfor ikke kun mere kommercielt værdifuldt end at eje opmærksomhed. Det kan også være farligere. De gamle platforme iagttog os for at udlede, hvem vi var. Den nye brugerflade inviterer os til at forklare os selv direkte.

Musk købte en ejendom, hvor andre mennesker hele tiden må skabe grunde til, at jeg kommer på besøg. OpenAI bygger måske en, hvor jeg selv bringer grunden med.

Musk forsøger at bygge Seahaven: stedet hvor alt sker. OpenAI bygger måske stedet, hvor alt begynder.

Beslægtede essays

AI & samfund · 16. august 2026AI åd internettet. Nu skal den afgøre, hvem der er menneskeVandmærkning skal gøre syntetisk indhold synligt. Den kan også gøre AI-selskaber til dem, vi beder om at attestere menneskeligt forfatterskab.

AI & edtech · 5. september 2026Når din kunde bliver din konkurrent: Kunden fandt en genvej til softwarefabrikkenFjerde del af serien om, hvordan 90 % af edtech forsvinder. AI gør vejen fra kontoret til softwarefabrikken kortere, og kunden behøver ikke længere købe det færdige produkt.

AI & edtech · 13. august 2026Edtech sender stadig dvd'er med postenHistorien om, hvorfor 90 % af edtech forsvinder: folk holdt aldrig op med at lære, de skiftede bare måde, og den forskel kan vælte store dele af branchen.

← Tilbage til blog

---

### https://sahra-josephine.com/da/blog/edtech-customer-competitor

← Blog

AI & edtech·5. september 2026Når din kunde bliver din konkurrent: Kunden fandt en genvej til softwarefabrikken

Denne tekst er maskinoversat fra engelsk. Nuancer kan gå tabt undervejs, så vi anbefaler at læse originalen. Læs den engelske original

Del fire i serien om, hvordan 90 % af edtech forsvinder. Softwarevirksomheder brugte tyve år på at overbevise kunderne om ikke at bygge selv. Kunstig intelligens ændrer afstanden mellem at ønske sig software og at lave det.

Jeg tilbragte en del af min sommerferie med at gå rundt i en af de smukkeste kommercielle fiaskoer, der nogensinde er bygget. Park Güell er i dag det bedste Barcelona: Gaudí, mosaikker, en ekstraordinær udsigt og en stabil strøm af turister, der fotograferer sig selv ved siden af en keramisk øgle. Men det var ikke tænkt som en offentlig park. Eusebi Güell og Antoni Gaudí planlagde det som et eksklusivt boligkvarter for velhavende familier, med tres huse placeret over byen. Kun to blev nogensinde bygget.

Vores officielle guide forklarede, at et af problemerne var overraskende praktisk. De mennesker, der var rige nok til at bo der, skulle stadig nå deres fabrikker og havnen, hvilket betød en tur med hestetrukken vogn ad en ubelejlig rute, hvor rejsen afhang af vejr og vejforhold. Kvarteret tilbød måske renere luft og smukke udsigter, men det var ganske enkelt for besværligt at komme fra huset til det sted, hvor pengene blev tjent.

Der var andre forhindringer. Grundstykkerne kom med restriktive betingelser, den offentlige transport var utilstrækkelig, og den eksklusivitet, der gjorde udviklingen attraktiv, bidrog også til at gøre den kommercielt uholdbar. Byggeriet stoppede i 1914. Güells arvinger solgte til sidst grunden til byen, og den åbnede som offentlig park i 1926.

I dag føles stedet ikke længere afsides. Barcelona voksede omkring det, mens veje, busser, taxaer og metroen ændrede den praktiske betydning af afstand. Park Güell var ikke nødvendigvis bygget det forkerte sted. Den var bygget, før vejen kom.

Jeg har tilbragt elleve år i SaaS i edtech-branchen med at fortælle kunder, at de ikke skulle bygge software, de kunne købe. Nioghalvfems procent af tiden var det oprigtigt fremragende råd. At bygge sin egen læringsplatform betød udviklere, designere, produktchefer, infrastruktur, integrationer, sikkerhed, vedligeholdelse og budget nok til at overleve det tidspunkt atten måneder senere, hvor nogen opdagede, at det interne system, alle havde arbejdet så hårdt på, var værre end det produkt, de kunne have købt fra starten.

Softwarefabrikken eksisterede, men for de fleste virksomheder var den langsom, besværlig og uhyre dyr at nå frem til. Så lærte vi virksomhederne at leje. Byg ikke dit eget learning management system. Køb et. Byg ikke et authoring-værktøj. Abonner på et. Byg ikke en medarbejderakademi- eller assessment-platform. Find en specialistvirksomhed, der allerede har løst problemet. Jeg byggede en succesfuld, venturefinansieret edtech-virksomhed på den logik.

I dag ejer jeg også et AI-studie, og jeg giver i stigende grad kunderne et råd, der ser ud til at være det modsatte: Før du køber endnu et softwareprodukt, find ud af, hvad det ville koste at bygge den funktionalitet, du faktisk har brug for. Det betyder ikke nødvendigvis at bygge det internt. En virksomhed kan spørge en af sine egne udviklere, en AI-assisteret konsulent eller et studie som mit. Den behøver ikke blive en softwarevirksomhed eller eje fabrikken. Den skal blot have overkommelig adgang til en.

Jeg ved, hvor omvæltende det råd er, fordi jeg selv har fulgt det. Jeg har tidligere skrevet om, at mit personlige nyhedsbrev kostede cirka 10.000 danske kroner om måneden at sende til omkring 36.000 mennesker gennem Mailchimp. Jeg brugte Lovable til at bygge mit eget afsendelsesværktøj og koblede det til SendGrid. Det kostede under 30 dollar at bygge, det koster cirka 25 til 30 dollar om måneden at drive, og den del, jeg ikke tidligere nævnte, er, at det tog under en dag.

Værktøjet er stadig ikke lige så godt som Mailchimp og mangler nogle af dets integrationer, skabeloner og edge cases. Men det har jeg ikke brug for. Det udfører den del af Mailchimps funktionalitet, jeg har brug for, og erstatter et abonnement, der kostede cirka 10.000 danske kroner (1.520 dollar) om måneden, eller omkring 18.000 dollar om året, med infrastruktur, der koster cirka 25 til 30 dollar om måneden, eller 300 til 360 dollar om året. Mailchimp tabte mig ikke til en anden email-platform. Det tabte mig til mig selv.

AI er ved at brolægge vejen mellem virksomhedens kontor og softwarefabrikken, og det er derfor, jeg i stigende grad fanger mig selv i at spørge kunder noget, jeg før ville have betragtet som forfærdeligt råd: Hvorfor køber du overhovedet det her?

Købet fandtes, før konkurrencen begyndte

De fleste softwarevirksomheder tænker på konkurrence, efter at kunden har besluttet sig for at købe. Kunden har brug for en læringsplatform, så den sammenligner læringsplatforme. Den har brug for et CRM, så den sammenligner CRM'er. Den har brug for et nyhedsbrevssystem, så Mailchimp konkurrerer med HubSpot, Klaviyo og hvad end ellers der optræder på shortlisten. Kategorien har allerede vundet; det eneste tilbageværende spørgsmål er, hvilken leverandør der får pengene.

Den antagelse ligger under en enorm mængde SaaS-strategi. Markedsføring skaber efterspørgsel efter kategorien, salg omsætter den efterspørgsel til en navngiven konto, produkt tilføjer nok funktioner til at slå den nærmeste konkurrent, og customer success beskytter fornyelsen. Selv churn forudsætter, at der først fandtes en kunde.

Tag en virksomhed med 10.000 medarbejdere og en elendig læringsopsætning. For ti år siden kunne HR have inviteret fem LMS-leverandører til at demonstrere deres produkter og valgt en. Spøg nu, hvad organisationen egentlig har brug for. Den har allerede et identity-system, politikker, videoer og intern viden. AI kan omdanne det materiale til øvelser. Det, der tilbage, er at gemme gennemførelse, rapportere til ledere, dokumentere overholdelse og forbinde delene.

Historisk set krævede selv den smalle arbejdsgang nok ingeniørarbejde til, at det stadig var rationelt at købe platformen. Nu kan virksomheden beskrive arbejdsgangen, forbinde sine eksisterende systemer og lade en intern produktperson, en AI-assisteret udvikler eller et eksternt studie bygge den del, den har brug for. Den kan stadig købe, hvis det kommercielle produkt er bedre, men at købe er ikke længere det automatiske udgangspunkt. Virksomheden behøver ikke genskabe LMS'et. Den skal genskabe grunden til, at den købte LMS'et.

Den farligste udsigt er derfor ikke kunden, der forsvinder ved fornyelse. I det mindste var den kunde i CRM'et, underskrev en kontrakt og skabte en grund til at forsvinde, som nogen kan analysere. Den mere urovækkende udsigt er virksomheden, der aldrig bliver en prospect: Den besøger aldrig prissiden, optræder aldrig i pipeline, anmoder aldrig om en demonstration eller giver indkøb en shortlist. Nogen spørger, om det her behøver at være endnu et abonnement, opdager, at det ikke behøver det, og tager den nye vej til softwarefabrikken i stedet. Fra SaaS-leverandørens synspunkt blev ingen handel tabt, for ingen handel fandtes nogensinde.

Du skal bare slå fakturaen

I størstedelen af SaaS-æraen kunne leverandører pakke hundredvis af funktioner sammen, fordi det at reproducere den delmængde, kunden brugte, var så dyr, at det gav mening at leje hele produktet. Den tærskel falder. Kunden behøver ikke bygge noget bedre, genskabe leverandørens samlede platform eller dække hver eneste edge case. Den skal bare løse sit eget problem godt nok til at slå fakturaen.

Tag en bevidst enkel læringsplatformspris på 12 dollar pr. bruger pr. måned, uden implementeringsgebyr, basisgebyr eller mængderabat. En virksomhed med 10.000 brugere betaler 1,44 millioner dollar om året. Ved 50.000 brugere er den årlige omkostning 7,2 millioner dollar; ved 200.000 er den 28,8 millioner. Det er ikke et påstand om, at en kompetent indkøbsafdeling ville acceptere den flade pris ved 200.000 sæder. Selv efter en 75 procents mængderabat ville det årlige abonnement dog stadig være 7,2 millioner dollar. Inputtene ændrer det punkt, hvor det bliver rationelt at bygge; de fjerner ikke forskellen mellem en omkostning, der stiger pr. bruger, og en der måske ikke gør.

Selvfølgelig ville en kunde af den størrelse forhandle. Enterprise-kontrakter er sjældent så rene, og en seriøs læringsplatform gør meget mere end at hoste nogle få sider og registrere gennemførelse. Den kan give hundredvis af integrationer, avancerede tilladelser, tilgængelighed, revisionsspor, indholdsstandarder, lokalisering, support, sikkerhedsdokumentation og kontraktlig ansvarlighed. En troværdig sammenligning må inkludere de ting.

Men det er netop pointen: Kunden behøver ikke reproducere alt, leverandøren har bygget. Forestil dig, rent som model, at en skræddersyet læringsplatform koster 1 million dollar at bygge og yderligere 2 dollar pr. bruger pr. måned at drive, sikre og vedligeholde. For 10.000 brugere ville første år koste cirka 1,24 millioner dollar, allerede under det 1,44 millioner dollar dyre abonnement. For 50.000 brugere ville det koste cirka 2,2 millioner dollar frem for 7,2 millioner; for 200.000 cirka 5,8 millioner frem for 28,8 millioner.

Det er illustrative tal, ikke en universel business case for at bygge. Regnestykket kan hurtigt vende, når en virksomhed har brug for global compliance, dusinvis af dybe integrationer, døgnåben support, komplekse migreringer eller en leverandør, der er villig til at påtage sig reel operationel risiko. Skræddersyet software skaber vedligeholdelses-, sikkerheds- og teknisk-gældsrisici, som et regneark kan skjule med imponerende effektivitet. Mange enterprise-platforme prissætter også efter aktive brugere, forbrug eller forhandlede bånd frem for at bruge en flad offentlig pris.

Selv efter generøse forbehold er retningen svær at ignorere. SaaS-priser pr. bruger stiger med kundens størrelse, mens omkostningen ved at bygge den relevante funktionalitet ikke nødvendigvis stiger i nær samme tempo. Større kunder betyder flere sæder og expansion revenue, men de kan også have den stærkeste økonomiske grund til at spørge, om leverandøren overhovedet skal eksistere i arbejdsgangen.

Ingen bruger seks måneder på at genskabe et produkt, der koster 200 euro om måneden. En lille skole bør nok ikke vedligeholde et hjemmelavet student-informationssystem, og en reguleret organisation bør passe på med følsomme data. En virksomhed, der betaler 250.000 euro om året, har et andet regnestykke. Din mest attraktive kunde kan også have den stærkeste motivation for aldrig at blive din kunde.

Det gamle spørgsmål var, om en custom build kunne måle sig med produktet. Det nye spørgsmål er, om den kan slå fakturaen, hvilket er en langt lavere tærskel.

Lejen rykker ned i stakken

Dette er ikke enden på at leje; det er en forandring af, hvad kunden lejer. Mit Mailchimp-erstatning bruger stadig SendGrid til at levere email. Jeg byggede ikke global email-infrastruktur, forhandlede direkte med hver eneste internetudbyder eller skabte mit eget system til at beskytte afsenderreputation. Jeg erstattede et applikationsabonnement med en tyndere samling af infrastruktur og et lille stykke software, jeg selv styrer.

En virksomhed, der bygger sit eget læringsmiljø, vil sandsynligvis gøre det samme. Den kan betale OpenAI, Anthropic eller Google for modeladgang, bruge AWS eller Azure til infrastruktur og beholde et eksternt team til at vedligeholde systemet. Softwarefabrikken er ikke blevet gratis. Den har stadig brug for elektricitet, maskiner, råvarer og nogen, der forstår, hvad der produceres, men lejen rykker ned i stakken.

Frem for at betale et stort løbende gebyr for en færdig applikation med 400 funktioner kan kunden betale mindre løbende gebyrer for infrastruktur og eje de syv funktioner, den faktisk bruger. Den kan bestille et one-time build uden at oprette en intern softwareafdeling, ligesom en modevirksomhed kan bestille produktion uden at eje hver eneste maskine, der laver dens tøj. Valget er ikke længere simpelthen byg eller køb. Det er køb, byg, samler eller bestil, og AI reducerer omkostningen ved de sidste tre.

SaaS vandt oprindeligt, fordi én specialistvirksomhed kunne bygge et produkt én gang og distribuere det billigt til tusindvis af kunder. Den fordel forsvinder ikke, men leverandøren skal dække omkostningen ved et generelt produkt, sit salgsapparat, sit customer-success-team, sine investorer og sin feature-roadmap over sin kundebase. Et custom system skal kun betjene én virksomhed. I årevar var det også dets svaghed, for én kunde kunne ikke retfærdiggøre rejsen til fabrikken. Nu er vejen kortere.

Edtech er særligt udsat, fordi to omkostninger falder på én gang. Generativ AI reducerer omkostningen ved at producere og koordinere læring: En leder kan bede en assistent om at forklare en politik, generere en øvelse, tilpasse materiale til en rolle, oversætte det og teste forståelse uden at åbne et kursuskatalog. Samtidig reducerer AI-assisteret udvikling omkostningen ved at bygge systemet, der tildeler, sporer og dokumenterer den læring. Den ene kraft angriber brugen; den anden angriber købet.

En edtech-virksomhed kan derfor blive klemt, selv mens efterspørgslen efter læring vokser. Organisationer vil stadig uddanne medarbejdere, dele viden, dokumentere compliance og udvikle kompetencer, og måske gøre mere af det hele. Men stigende efterspørgsel efter resultatet garanterer ikke stigende efterspørgsel efter den eksisterende produktkategori. Folk holdt ikke op med at se film, da DVD-udlejning kollapsede, eller med at lytte til musik, da det blev absurd at købe CD'er. Aktiviteten overlevede; produktet, distributionen og betalingsmodellen omkring den ændrede sig.

Læring forsvinder ikke. Antagelsen om, at den skal pakkes ind i en separat købt læringsplatform, bliver mindre sikker.

Måske var din voldgrav afstanden

Den indlysende indvending er, at vibe-kodet software er upålideligt, usikkert og let at demonstrere, men svært at drive. Ofte er det sandt. Der er en enorm afstand mellem at lave en fungerende prototype og at køre et business-kritisk system, som har brug for arkitektur, tests, tilladelser, overvågning, backups, data governance, tilgængelighed, sikkerhedsgennemgang og nogen, der står til ansvar, når det går i stykker. AI kan producere dårlig kode meget hurtigt og lade en virksomhed skabe en skrøbelig intern afhængighed, som ingen forstår seks måneder senere.

Men svag eksekvering redder ikke en svag forretningsmodel; den betyder bare, at kunderne har brug for en kompetent vej til at bygge. Det tidlige internet var fuldt af forfærdelige hjemmesider, og det reddede hverken avisernes rubrikannoncer, rejsebureauerne eller detailhandlen. Dårlige første forsøg kan sameksistere med en strukturel forandring i omkostning og adgang.

Det mere nyttige spørgsmål er, hvilken del af en SaaS-virksomheds forsvar der kom fra ægte vanskeligt arbejde, og hvilken del der kom fra, at kunden var for langt fra fabrikken. Dybe integrationer, proprietære data, regulatorisk godkendelse, kontraktlig ansvarlighed, betroede credentials, distribution, fællesskab og demonstrerede resultater kan være formidabelt. En leverandør, der forstår et kompliceret domæne og påtager sig ansvaret for at drive det, kan være langt mere værd end sine funktioner.

At have mange funktioner er ikke nødvendigvis en voldgrav. En smuk brugerflade er mindre holdbar, når brugerflader kan genereres, et årtis akkumuleret kode betyder mindre, når kunden kun har brug for en smal arbejdsgang, og switching costs beskytter kun begrænset mod en virksomhed, der endnu ikke har købt noget. Mange SaaS-virksomheder var sikre, delvis fordi selv en dårligere version krævede et team, kunden ikke havde, og et budget, den ikke kunne retfærdiggøre.

AI gør ikke hvert produkt let, sikkert eller fornuftigt at genopbygge. Den gør nok produkter billige nok til at sætte spørgsmålstegn ved. En hårdere definition af forsvar følger: En voldgrav er ikke det, der gør dit produkt svært at genskabe. Det er det, der gør kundens køb svært at fjerne.

I tyve år svarede softwarepositionering mest på hvorfor os frem for dem? Kunden havde allerede accepteret, at den havde brug for at købe noget, og leverandørens opgave var at vinde sammenligningen. I stigende grad bliver softwarevirksomheder nødt til at svare på et farligere spørgsmål: Hvorfor købe overhovedet?

Det er et meget sværere argument, fordi virksomheden ikke indvender mod din pris, beder om endnu en funktion eller truer med at vælge din konkurrent. Den kommer måske aldrig ind på dit marked. Ingen mulighed optræder i CRM'et, ingen indkøbsproces begynder, og ingen lost-deal-analyse forklarer, hvad der skete. Behovet bliver simpelthen løst gennem software, virksomheden ejer, frem for software, den lejer.

Funktionaliteten forbliver. Medarbejdere lærer stadig, ledere har stadig brug for information, og compliance har stadig brug for beviser. Det, der forsvinder, er transaktionen i midten.

Park Güell var ikke bygget det forkerte sted. Den var bygget, før infrastrukturen ændrede, hvad det sted betød. SaaS blev bygget til en verden, hvor vejen til softwarefabrikken var for lang og dyr for de fleste virksomheder at tilbagelægge, så den omdannede fabrikkens output til noget, de kunne leje. Det var oprigtigt godt råd, indtil vejen ændrede sig.

Softwarefabrikken er ikke forsvundet, og heller ikke behovet for det, den laver. Men virksomheder kan i stigende grad nå den uden at passere den SaaS-leverandør, der forventede at sælge dem et abonnement, hvilket betyder, at din konkurrent ikke er et andet produkt. Det er forsvindingen af købet.

Serie · Sådan forsvinder 90 % af edtech

En serie i fire dele om pengene, markedet og modellerne, der gjorde en hel branche til noget andet.

Del 1 af 4Hvis jeg drev en edtech-fond, ville jeg skide i bukserne

Del 2 af 4Din edtech-investor vil have et åbent forhold

Del 3 af 4 · Forrige i serienEdtech sender stadig dvd'er med posten

Del 4 af 4Når din kunde bliver din konkurrent: Kunden fandt en genvej til softwarefabrikken

Læs hele serien →Beslægtede essays

AI & edtech · 18. juli 2026Hvis jeg ser én til pushe “PedTech”, kaster jeg opPædagogikken har aldrig manglet som idé. Det svære har været at bygge en branche, hvor økonomien lod pædagogikken blive stående i centrum, når investorerne bankede på.

AI & samfund · 8. september 2026Elon Musk købte den forkerte ejendom til alt-appenMusk købte Twitter for at bygge alt-appen. Men mens han samlede opmærksomhed, samlede OpenAI intentioner, og det kan være den mest værdifulde adresse på internettet.

AI & samfund · 16. august 2026AI åd internettet. Nu skal den afgøre, hvem der er menneskeVandmærkning skal gøre syntetisk indhold synligt. Den kan også gøre AI-selskaber til dem, vi beder om at attestere menneskeligt forfatterskab.

← Tilbage til blog

---

### https://sahra-josephine.com/da/blog/ai-human

← Blog

AI & samfund·16. august 2026AI åd internettet. Nu skal den afgøre, hvem der er menneske

Denne tekst er maskinoversat fra engelsk. Nuancer kan gå tabt undervejs, så vi anbefaler at læse originalen. Læs den engelske original

Vandmærkning skal gøre syntetisk indhold gennemsigtigt. Den kan også gøre AI-selskaber til de institutioner, vi beder om at attestere menneskeligt forfatterskab.

Da jeg var barn, skrev jeg mine skoleopgaver i hånden og maskinskrev dem bagefter på vores skrivemaskine.

Jeg skulle egentlig selv stå for maskinskrivningen. Det var en nyttig ting at lære. Men nogle gange blev det sent, opgaven var allerede skrevet, jeg var udkørt, og min mor overtog, så jeg kunne komme i seng. Hun var enlig mor og havde bedre ting at bruge tiden på end at holde øje med et overtræt barn, der højtideligt hamrede de sidste taster ned klokken 23.

Ingen mente derfor, at hun havde skrevet min opgave. Ingen kaldte det snyd, plagiat eller en falsk aflevering. Vi forstod intuitivt, at der var forskel på den evne, opgaven skulle vise, og den infrastruktur, der blev brugt til at fremstille det færdige produkt. Ideerne, argumentet, sproget og forståelsen var mine. At kunne maskinskrive havde en vis værdi som færdighed, men det var ikke det, opgaven handlede om.

Så fik vi en computer. Det var enormt spændende, ikke mindst fordi den samme maskine, jeg pludselig kunne skrive mine opgaver på, også lod mig spille Gorillas, det herligt fjollede spil, hvor to gorillaer kastede eksploderende bananer efter hinanden hen over hustagene.

Nu leverede computeren selv skriveinfrastrukturen. Min mor blev til stavekontrollen, læseren og den, jeg vendte ideer med. Så overtog softwaren rollen som stavekontrol. Grammatikprogrammer begyndte at rette sætninger. Google blev en del af researchprocessen. Autofuldførelse begyndte at gætte, hvad jeg var ved at skrive.

Ingen krævede en kvittering for præcis, hvilke ord min mor, Microsoft Word eller Google havde rørt ved.

Og så kom AI'en.

Hvorfor besluttede vi lige netop her, at infrastruktur bliver til forfatterskab?

Jeg bruger AI til en blanding af alle de roller, jeg lige har nævnt: maskinskriver, researcher, redaktør, kritiker, stavekontrol og sparringspartner. Noget af det er infrastruktur. Noget er intellektuel bistand. Nogle gange er det reelt intellektuelt arbejde.

Men intellektuelt arbejde og forfatterskab har aldrig været det samme. Min mor bidrog intellektuelt, når jeg vendte ideer med hende. En redaktør kan forbedre et argument betragteligt. En researcher kan finde den kendsgerning, hele essayet hviler på. Ingen af dem bliver automatisk forfatteren af den grund.

Forfatterskab handler nok snarere om, hvor den styrende intellektuelle kraft sidder: hvem der bestemmer, hvad værket handler om, vælger mellem ideer, forstår argumentet, afgør hvad der hører hjemme i det, og tager ansvar for resultatet.

For år tilbage eksperimenterede vi hos CanopyLAB med at bruge kunstig intelligens til at vurdere elevopgaver på næsten den modsatte måde af, hvad vandmærkning gør i dag. Én samlet karakter var for groft et redskab. Vi ville skelne mellem ting som forståelse, argumentationsstyrke, sammenhæng, grammatik og stil, og bruge de skel til at give bedre, kvalitativ feedback.

Dengang gjorde teknologien det svært. Så kom generativ AI, og størstedelen af det tekniske problem blev nærmest irrelevant. Moderne modeller kan se på en tekst og adskille de forskellige præstationsdimensioner på få sekunder. Så jeg tror ikke, problemet er, at vi aldrig kunne finde ud af, hvad Claude bidrog med. Vi kunne sandsynligvis vide langt mere, end vi gør.

Anthropic siger, at fremtidige Claude-modeller vil vandmærke tekst som led i selskabets svar på EU's krav om gennemsigtighed. I dag kan vandmærket vise, at Claude har været involveret, uden nødvendigvis at skelne mellem, om Claude har genereret noget fra bunden, eller blot redigeret kraftigt i det. Men der er ingen grund til at tro, at proveniens skal forblive så groft et redskab for evigt.

Måske bliver fremtidens kvittering langt mere detaljeret. Claude foreslog strukturen. Claude fandt modargumentet. Claude omskrev fire afsnit. Claude rettede grammatikken. Mennesket forkastede seks forslag, tjekkede kilderne efter, omskrev halvdelen af den genererede tekst og traf de endelige beslutninger.

Fint nok. Så har vi den fulde ingrediensliste. Vi har stadig ikke svaret på, hvem der lavede maden. Det er den forskel, jeg interesserer mig for. Bedre måling af bidrag giver os ikke automatisk en teori om forfatterskab.

Ægthedsministeriet

Anthropic gør ikke krav på at afgøre, hvem der har forfattet et værk. Det mere interessante er, at alle andre ser ud til selv at give dem den rolle.

Et universitet vil vide, om en studerende har brugt Claude. Et forlag vil håndhæve en AI-politik. En arbejdsgiver vil tjekke en ansøgning. En litterær pris skal afgøre, hvor meget AI-hjælp den vil acceptere. Snart findes der et API, der kan svare på et spørgsmål, disse institutioner ikke selv kan besvare: var Claude sandsynligvis med her?

Ingen hos Anthropic behøver at udråbe et Ægthedsministerium. Alligevel bygger vi et op omkring dem selv. Og det er her, konsekvenserne bliver mere interessante end spørgsmålet om, hvorvidt vandmærket virker.

Forestil dig, at en studerende skriver en opgave, bruger Claude til at udfordre argumentet og stramme sproget, og at vandmærket korrekt afslører Claudes involvering. Detektoren har ikke begået nogen fejl. Men en skole kan sagtens begå en, hvis den tolker det korrekte signal som bevis for, at den studerende ikke selv har lavet det intellektuelle arbejde.

Det samme kan ske for en forfatter, der bruger AI som redaktør, en jobsøger, der bruger den til at rette akavede formuleringer, eller en forsker, der skriver på et fremmedsprog og bruger den til at polere en fondsansøgning.

Den farligste udgave af dette system er ikke én, der fejlagtigt påviser AI. Det er én, der korrekt påviser AI-deltagelse og derefter drager en helt uberettiget konklusion om mennesket bag.

Og incitamenter gør, hvad incitamenter altid gør. Bliver det strafbart at have kvitteringen liggende, lærer folk sig at destruere den. De, der bruger AI åbent og legitimt, bliver dem, det er lettest at afsløre, mens dem, der faktisk prøver at skjule dens rolle, får al mulig grund til at fjerne vandmærket, “humanisere” teksten eller skifte til værktøjer, der ikke efterlader spor.

Vi risikerer altså at indføre proveniens for at skabe gennemsigtighed og ende med at belønne dem, der skjuler det i stedet. Teknologien kan virke præcis, som den skal. Problemet opstår, når vi belønner og straffer folk ud fra, hvad vi selv beslutter, at kvitteringen betyder.

Infrastruktur er ikke kompetence

I undervisning er forskellen indlysende. Vil jeg teste hovedregning, tager jeg lommeregneren fra eleven. Vil jeg teste maskinskrivning, ødelægger det formålet, hvis min mor færdiggør min. Men vil jeg vide, om nogen kan analysere beviser, opbygge et argument og forsvare en konklusion, skal jeg vurdere netop det.

Spørg, hvor modellen tog fejl. Spørg, hvad den studerende accepterede, og hvad de forkastede. Spørg, hvorfor de valgte det ene argument frem for det andet. Bed dem forsvare det, de har afleveret, og tage ansvar for det.

“Har AI været inde over det her?” er nemmere at måle. Det er også et andet spørgsmål.

Før AI var det færdige produkt en nogenlunde brugbar indikator for det hoved, der havde produceret det. En studerende afleverede et essay, og vi udledte noget om deres forståelse ud fra essayet. AI ødelægger den indikator. Men svaret behøver ikke at være at genopbygge den gennem stadig mere avanceret kriminalteknisk analyse af produktet. Vi kan vurdere tankearbejdet mere direkte.

Og vi ved allerede, at vi kan gøre det bedre. Hele løftet ved at bruge AI i vurdering var, at vi kunne komme ud over de grove signaler og forstå mere om mennesket bag: hvad de forstod, hvor deres ræsonnement var stærkt, hvad de kæmpede med, og hvad de skulle arbejde videre med.

Pointen var at give bedre og mere nuanceret feedback. Det ville være en mærkelig kovending, hvis langt mere kapabel AI nu hjælper os med at reducere den samme menneskelige præstation til en kriminalteknisk etiket for, hvilket værktøj der var involveret.

Værktøjer findes, fordi alternativomkostninger findes. Jeg kunne regne et regneark ud i hånden, men Excel findes. Jeg kunne slå en statistik op i tyve bøger i stedet for at finde den på minutter. Jeg kunne bruge tre timer på mekanisk at skære gentagelser væk fra en artikel, eller bruge noget af den tid på at udvikle det næste argument.

At jeg selv kunne udføre den opgave, der giver mindst værdi, gør det ikke intellektuelt fornemmere at gøre det selv.

Meget teknologisk fremskridt handler om at gøre dyre eller tidskrævende opgaver til infrastruktur og dermed frigøre menneskelig opmærksomhed til noget andet. AI kan meget vel være det næste lag af intellektuel infrastruktur.

Det betyder ikke, at forståelse bliver mindre vigtig. Jeg tror faktisk, det modsatte gør sig gældende. Jo mere udførelse vi kan outsource, jo vigtigere bliver dømmekraft. Kan en maskine levere et plausibelt svar på få sekunder, bliver det langt mere værdifuldt at kunne vurdere, om svaret er godt, dårligt, opdigtet, afledt eller irrelevant.

Færdigheden flytter sig. Vores vurdering bør flytte sig med.

Og på en eller anden måde ender maskinen med den bedste kvittering

Generativ AI blev nyttig ved at lære fra et enormt hav af menneskelig intellektuel produktion. Forfattere skrev. Forskere forskede. Programmører kodede. Lærere forklarede. Millioner af mennesker skabte det informationsmiljø, som gjorde disse systemer værdifulde.

Der er uafklarede juridiske spørgsmål om træningsdata, men det er ikke min pointe her. Det, jeg interesserer mig for, er retningen på tilskrivningen.

Menneskelig viden strømmer ind i AI-systemerne i enorm skala, og det enkelte bidrag bliver umuligt at udpege. Så deltager maskinen i vores næste stykke arbejde, og pludselig bliver dens bidrag vigtigt nok til, at vi gør det teknisk sporbart.

Menneskeheden leverer den intellektuelle råvare. Maskinen får kvitteringen.

Gør den kvittering nok så detaljeret. Fortæl mig præcis, hvilken sætning Claude foreslog, hvilken kilde den fandt, hvilket afsnit den omstrukturerede, og hvilket komma den flyttede. Jeg tror stadig ikke, at maskinens proveniens bør blive den målestok, en anden bruger til at afgøre, om værket er mit.

Vi forstod forskellen dengang, et udmattet barn gik i seng, og hendes mor færdiggjorde et essay, der allerede var skrevet. Ingen havde brug for en kriminalteknisk logbog over, hvor mange taster min mor trykkede på, eller hvilke stavekorrektioner der kom fra hende. Vi brugte sund fornuft, fordi vi forstod, hvad opgaven i virkeligheden skulle måle.

Måske skulle vi huske det, inden vi bygger en ægthedsinfrastruktur op omkring det, der er lettest at opdage.

Kvitteringen fortæller os, at maskinen var der. Den kan stadig ikke fortælle os, hvem der tænkte.

Beslægtede essays

AI & samfund · 8. september 2026Elon Musk købte den forkerte ejendom til alt-appenMusk købte Twitter for at bygge alt-appen. Men mens han samlede opmærksomhed, samlede OpenAI intentioner, og det kan være den mest værdifulde adresse på internettet.

AI & edtech · 5. september 2026Når din kunde bliver din konkurrent: Kunden fandt en genvej til softwarefabrikkenFjerde del af serien om, hvordan 90 % af edtech forsvinder. AI gør vejen fra kontoret til softwarefabrikken kortere, og kunden behøver ikke længere købe det færdige produkt.

AI & edtech · 13. august 2026Edtech sender stadig dvd'er med postenHistorien om, hvorfor 90 % af edtech forsvinder: folk holdt aldrig op med at lære, de skiftede bare måde, og den forskel kan vælte store dele af branchen.

← Tilbage til bloggenLæs mere fra bloggenOplægInviter mig til at holde oplæg

---

### https://sahra-josephine.com/da/blog/edtech-dvds

← Blog

AI & Edtech·13. august 2026Edtech sender stadig dvd'er med posten

Historien om, hvorfor 90 % af edtech forsvinder. Folk er ikke holdt op med at lære, de har bare fundet en ny måde at gøre det på. Den forskel kan udslette store dele af edtech-branchen, som vi kender den.

Denne tekst er maskinoversat fra engelsk. Nuancer kan gå tabt undervejs, så vi anbefaler at læse originalen. Læs den engelske original

Da jeg gik på universitetet, var Netflix et firma, der sendte dvd'er med posten. Man lavede en liste over alt, man gerne ville se, en “kø”, som Netflix kaldte den. Standardabonnementet i 2002 gav dig tre dvd'er ude ad gangen, uden afleveringsfrist og uden gebyrer for at komme for sent, og der fandtes også andre abonnementer med flere eller færre dvd'er ad gangen. Sendte man én tilbage, sendte Netflix straks den næste film fra ens kø.

Dengang var det genialt. Man skulle ikke ud og køre til Blockbuster. Der var ingen gebyrer. Kataloget var enormt sammenlignet med, hvad en fysisk butik kunne have på hylderne, og Netflix lærte løbende, hvad man kunne lide, så de kunne anbefale, hvad man skulle se næste gang. Men der var én kæmpe begrænsning, som Netflix aldrig kunne optimere sig ud af: man skulle stadig vente på, at posten kom med den næste dvd.

Da Netflix lancerede streaming i 2007, skrottede de gradvist den fysiske levering og bredte i stedet distributionen ud fra computeren til alle mulige andre enheder. Kunderelationen bestod. Anbefalingssystemet bestod. Det meste af det, der egentlig skabte værdien, bestod. Det var kun den røde konvolut og ventetiden, der forsvandt.

Jeg tænker mere og mere på edtech ud fra præcis den skelnen. I mit første essay skrev jeg, at op mod 90 % af nutidens edtech vil forsvinde, blive slugt af noget andet eller ganske enkelt gøre sig selv irrelevant i de kommende år. Det er ikke, fordi uddannelse betyder mindre. Det er, fordi læring lige nu oplever sit eget Netflix-øjeblik.

Folk søger stadig information. De prøver stadig at forstå svære begreber, øve færdigheder, læse op til eksamen, lære sprog, skifte karriere, løse problemer og få ting forklaret. Det, der har ændret sig radikalt, er hvordan.

De åbner ChatGPT, Claude, Gemini eller Perplexity. De spørger direkte og følger op med endnu et spørgsmål. De uploader det dokument, de sidder fast i. De beder om en forklaring, en tolvårig kan følge, og bagefter én på kandidatniveau. De beder om eksempler, øvelser og modspørgsmål, diskuterer svaret og beder systemet om at forklare, hvor de tog fejl.

Det er ikke længere gætværk. OpenAI oplyste i 2025, at mere end en tredjedel af de unge i universitetsalderen i USA brugte ChatGPT, og at omkring en fjerdedel af deres beskeder handlede om læring eller skolearbejde. UNESCO rapporterede, at over to tredjedele af eleverne på ungdomsuddannelser i højindkomstlande allerede brugte generativ AI, mens Anthropics egne brugsdata fra 2026 pegede på undervisning og litteraturarbejde som den næststørste brugskategori på Claude.ai.

Læring er ikke forsvundet. Det er måden, vi konsumerer den på, der har ændret sig. Og en stor del af edtech sender stadig dvd'er med posten.

Edtechs ubehagelige forhistorie

Den nemme forklaring på, hvorfor så meget edtech pludselig ser skrøbeligt ud, er generativ AI. Det ville også være bekvemt at sige, at ChatGPT kom, og at en ellers sund branche på et øjeblik blev forældet. Problemet er bare, at tidslinjen slet ikke passer.

Branchen havde alvorlige strukturelle problemer længe før ChatGPT. Kigger man på de sidste 10-15 år af edtech, kan man næsten følge en lang række forsøg på at rydde op efter den forrige generations fejl. De tidlige virksomhedssystemer var som regel elendige. Svære at købe, svære at sætte op, svære at bruge. Det kunne kræve konsulenter, integrationer, oplæring og måneders arbejde bare at få det op at køre, før en eneste studerende havde set skyggen af det.

Så kom SaaS og lignede løsningen. Produkterne blev lettere at købe, lettere at implementere og bedre designet. Cloud, abonnement, intet stort it-projekt. Det var reelt et fremskridt, men på et tidspunkt begyndte branchen at forveksle mindre friktion i salget med mere værdi for kunden.

Vi blev ekstremt gode til at lave software: en kursusplatform, en quizplatform, en tutorplatform, en lærerplatform, en engagementsplatform, en sprogapp, en trivselsapp, en app til privatøkonomi, en app til kompetenceudvikling, endnu en platform til endnu en niche. Endnu et login, endnu et dashboard, endnu et abonnement. Fordi SaaS gjorde det let at sælge software, blev der solgt enorme mængder software. Det betyder bare ikke, at nogen brugte det.

UNESCO's Global Education Monitoring Report fra 2023 viste, at i gennemsnit 67 % af licenserne til undervisningssoftware i USA slet ikke blev brugt, og at 98 % ikke blev brugt intensivt. EdTech Genome Project, som rapporten citerer, gennemgik omkring 7.000 undervisningsværktøjer for i alt 13 milliarder dollars og fandt, at 85 % passede dårligt til formålet eller var dårligt implementeret. Det er et opsigtsvækkende tal, og branchen har efter min mening haft det alt for let ved bare at kalde det et implementeringsproblem.

Implementering betyder selvfølgelig noget. Skolerne uddanner ikke lærerne ordentligt. Ledelsen køber software uden at ændre praksis. Produkter integreres dårligt. Lærerne har ikke tid. Indkøb og faktisk brug lever hver sit liv. Men set ud fra forretningsmodellens logik rækker de forklaringer kun et stykke.

Hvis jeg bygger et produkt, der kun skaber værdi, når kunden uddanner alle grundigt, integrerer det perfekt, holder brugen oppe og på en eller anden måde overtaler tusindvis af travle mennesker til at tage en ny vane til sig, så er de forudsætninger en del af mit produkts økonomi. Jeg kan ikke bare skyde skylden over på kunden.

Evidensen var lige så ubehagelig. UNESCO fandt, at kun 7 % af de britiske edtech-virksomheder i deres undersøgelse havde lavet et randomiseret kontrolleret forsøg. 12 % havde brugt tredjepartscertificering, 18 % havde deltaget i akademiske studier. I en undersøgelse på tværs af 17 amerikanske delstater bad kun 11 % af lærere og skoleledere om peer reviewet evidens, før de indførte ny undervisningsteknologi.

Resultatet blev et besynderligt marked. Virksomhederne dokumenterede sjældent, at deres produkter forbedrede læring. Køberne krævede sjældent dokumentation, før de købte dem. Og branchen blev ekstremt dygtig til at måle alt det, software nu engang kan måle: oprettelser, logins, sete minutter, besvarede spørgsmål, gennemførte kurser, udstedte beviser.

Den slags tal fortæller, hvad der skete inde i softwaren. De fortæller ikke nødvendigvis, at nogen lærte noget, og de fortæller bestemt ikke, at kunden fik nok værdi til at blive ved med at betale.

Covid fik næsten alle dvd'er til at se nyttige ud, i en periode

Så kom covid. Den globale venturekapital i edtech lå på omkring 7 milliarder dollars i 2019. Den sprang til 16,1 milliarder i 2020 og 20,8 milliarder i 2021. I 2024 var den styrtdykket til omkring 2,4 milliarder, før den kravlede lidt op igen til 2,6 milliarder i 2025.

Den gængse forklaring er, at investorerne mistede besindelsen under pandemien og siden kom til fornuft. Der er noget om det, men covid skabte også en ekstraordinær illusion af product-market fit, simpelthen fordi alternativet forsvandt.

Skolerne valgte ikke frivilligt fjernundervisning frem for klasselokaler, der fungerede fint. Universiteterne kunne ikke bare vente på, at softwaren blev bedre. Virksomhederne kunne ikke udskyde al oplæring på ubestemt tid, og familierne kunne ikke beslutte, at nu havde børnene fået skærmtid nok, og bare sende dem tilbage i skole. Der var ingen andre steder at gå hen.

Det mærkede jeg på egen krop. Før pandemien kunne virksomheder som min bruge halvdelen af en salgssamtale på at forklare, hvorfor læring overhovedet burde foregå online. Pludselig behøvede vi ikke det længere. Verden havde ført argumentet for os.

Brugen eksploderede, men nødadoption viser kun, at et produkt fungerer, når folk ingen andre muligheder har. Det viser ikke, at de bliver ved med at vælge det, når muligheden kommer tilbage. Branchen behandlede den nødtrafik, som om et helt nyt adfærdsmønster pludselig var blevet permanent. Så åbnede verden op igen, og kunderne fik deres valg tilbage.

Hvis covid bare midlertidigt maskerede strukturelle svagheder, der allerede var der, burde de svagheder være dukket op, inden generativ AI blev et seriøst alternativ. Det oplagte sted at lede efter beviset er det børsnoterede marked.

Markedet knækkede, før ChatGPT kom

Jeg ville vide, om de svagheder, covid midlertidigt havde skjult, viste sig, før generativ AI blev et reelt alternativ. Ikke fordi en aktiekurs kan sige noget om, hvorvidt folk gerne ville lære noget, det kan den ikke, men fordi den kan sige noget langt snævrere og alligevel vigtigt: hvad der skete med de penge, der blev sat i de virksomheder, som skulle profitere på digitaliseringen af uddannelse.

Så jeg gik tilbage til udgangen af 2020 og tog de 20 største rendyrkede børsnoterede uddannelsesvirksomheder efter markedsværdi på det tidspunkt: TAL Education Group, Offcn Education, New Oriental, Chegg, Gaotu Techedu, Bright Horizons, Pearson, Kahoot!, Grand Canyon Education, China Education Group, China East Education, IDP Education, East Buy Holding, China Yuhua Education, Youdao, Laureate Education, Pluralsight, 2U, Cornerstone OnDemand og John Wiley & Sons.

Så lavede jeg et bevidst simpelt eksperiment. Forestil dig, at du den 31. december 2020 havde en million dollars og besluttede, at edtech var stedet at putte dem. Hvad var der sket med de penge?

Øvelsen har åbenlyse begrænsninger. En aktiekurs måler ikke pædagogik, læringsbehov eller produktkvalitet. En virksomhed kan vokse i omsætning og alligevel ødelægge værdi for sine aktionærer, fordi investorerne betalte for meget fra starten. En anden kan bygge et fantastisk produkt og alligevel give et middelmådigt afkast. Kurser afhænger af renter, eksekvering, konkurrence, regulering, geografi og de forventninger, der allerede lå indbygget i den oprindelige værdiansættelse.

Så jeg bruger ikke aktiemarkedet som bevis for, at folk holdt op med at ville lære. Jeg bruger det til at stille et langt snævrere spørgsmål: blev dem, der satte kapital i de virksomheder, som repræsenterede kategorien, rent faktisk belønnet for det?

Der var også en stor komplikation blandt de oprindelige 20. Ni var kinesiske, og de restriktioner på privatundervisning, Kina indførte i 2021, ændrede fuldstændig økonomien i en stor del af det private uddannelsesmarked. De tab er helt reelle, hvis man var investor, men de fortæller mindre om, hvorvidt forretningsmodellerne allerede var ved at smuldre af sig selv, uafhængigt af en ekstraordinær regulatorisk begivenhed. Så jeg regnede på begge dele.

Den fulde portefølje på 20 virksomheder viser, hvad der faktisk skete, hvis man købte hele kategorien, som den så ud ved udgangen af 2020, Kina inklusive. Til den centrale sammenligning fjernede jeg de ni kinesiske virksomheder og kiggede på de resterende 11: Chegg, Bright Horizons, Pearson, Kahoot!, Grand Canyon Education, IDP Education, Laureate Education, Pluralsight, 2U, Cornerstone OnDemand og John Wiley & Sons.

Og så spurgte jeg: hvad skete der med den million dollars?

I gruppen uden Kina blev en million dollars investeret i medianvirksomheden til omkring 1,07 millioner. Fordelt ligeligt på alle 11 virksomheder ville de være blevet til omkring 1,31 millioner, primært fordi en håndfuld stærke virksomheder trak resten af porteføljen med op. Var millionen i stedet fordelt efter virksomhedernes markedsværdi ved periodens start, altså efter det hierarki markedet selv allerede havde givet kategorien, ville den være endt på omkring 1,02 millioner.

I samme periode blev en million dollars i SPY til omkring 2,10 millioner. I QQQ til omkring 2,18 millioner. Den sammenligning er svær at pynte på.

Den, der ville investere specialiseret i edtech, skulle forstå skoleindkøb, pædagogik, implementering, gennemførelsesprocenter, regulering, institutionelle budgetter, B2B-salgscyklusser og forskellen på et LMS, en OPM, en tutorvirksomhed og en universitetsoperatør. Vedkommende skulle vurdere, hvilke virksomheder der fortjente kapital, hvilke værdiansættelser der gav mening, og hvilke forretningsmodeller der ville overleve. Den passive investor kunne bare købe SPY og gå til frokost.

Fulgte man markedets oprindelige hierarki blandt uddannelsesvirksomhederne uden Kina, sad specialisten tilbage med omkring 1,02 millioner. Det samme beløb i SPY gav omkring 2,10 millioner. Mere end en million dollars i ekstra værdi kom simpelthen af at vælge det kedelige, passive alternativ.

Det beviser ikke, at edtech var nytteløst, og det beviser ikke, at der ikke var efterspørgsel efter digital uddannelse. Det beviser noget langt mere ubehageligt for investorerne: som gruppe belønnede de største børsnoterede uddannelsesvirksomheder ikke den ekstra specialisering, koncentration og risiko, det krævede at investere i dem.

Men selv den konklusion er for grov, for virksomhederne opførte sig slet ikke ens.

En million dollars i Laureate Education blev til omkring 5,67 millioner. I Pearson til omkring 2,21 millioner. Grand Canyon Education nåede omkring 1,71 millioner, Wiley omkring 1,45 millioner og Cornerstone OnDemand omkring 1,31 millioner.

I den anden ende faldt en million dollars i Chegg til omkring 12.000 dollars. I 2U gik den helt til nul. IDP Education faldt til omkring 122.000, Kahoot! til omkring 368.000 og Bright Horizons til omkring 478.000.

Det er ikke små udsving omkring et fælles branchegennemsnit. Én virksomhed mere end femdoblede den oprindelige investering, mens en anden stort set udslettede den.

Og det betyder noget, for det udelukker den letteste konklusion. Markedet besluttede ikke, at uddannelse i sig selv var blevet værdiløst. Det skelnede mellem vidt forskellige typer forretninger, der bare var havnet under samme edtech-mærkat.

Måske ligger en del af problemet i selve kategorien. Vi kalder en enorm vifte af fundamentalt forskellige forretninger for “edtech”, blot fordi de alle rører ved uddannelse. Men en universitetsoperatør, et forlag, en testleverandør, et LMS, en tutorplatform og en læringsapp til forbrugere deler ikke nødvendigvis samme økonomi, bare fordi kunderne lærer noget.

Nogle af de stærkeste forretninger var bundet til institutioner, eksamensbeviser, prøvning, forskning, udgivelse, regulerede aktiviteter eller infrastruktur, der sad dybt inde hos kunden. Andre afhang langt mere af, at folk igen og igen valgte at besøge et separat digitalt sted for indhold, kurser, aktiviteter eller svar. Og det var i den sidste gruppe, den mest spektakulære værdiødelæggelse fandt sted.

Det leder til det næste oplagte spørgsmål. Hvis de forretninger var særligt sårbare, fordi folk skulle vende tilbage til en separat platform for noget, de i stigende grad kunne finde andre steder, var det så ChatGPT, der knækkede dem?

Tidslinjen siger nej.

Det første fald kom før AI. Så kom endnu et.

Chegg er det mest ekstreme eksempel. På toppen i 2021 var virksomheden omkring 14,7 milliarder dollars værd. I dag ligger markedsværdien på omkring 100 millioner. Mere end 99 % af værdien er væk.

Det er ekstremt fristende at kigge på den ødelæggelse og fortælle den enkle historie: ChatGPT kom, de studerende holdt op med at betale for svar, og en hel kategori brød sammen. Problemet er bare, at det første brud kom for tidligt til den forklaring.

Jeg delte de børsnoterede uddannelsesvirksomheder op efter, om en generel AI kunne levere det samme centrale resultat, som deres kunder betalte for. Det handlede ikke om, hvorvidt virksomhederne selv brugte AI. Det handlede om, hvor let et generelt AI-system kunne erstatte det, kunden tidligere skulle bruge virksomheden til.

Fra februar 2020 til november 2022 faldt en million dollars i medianvirksomheden med høj eksponering til omkring 550.000. I samme periode blev en million i medianuddannelsesforretningen med lav eksponering til omkring 1,39 millioner.

Det skete, før ChatGPT overhovedet havde en chance for at være årsagen. Noget var allerede galt.

Så gjorde mønsteret noget vigtigt, man let overser, hvis man kun kigger på start og slut: det stabiliserede sig. Fra november 2022 til december 2023 voksede en million dollars i medianvirksomheden med høj eksponering til omkring 1,04 millioner, mens den samme investering i gruppen med lav eksponering nåede omkring 1,06 millioner. I omtrent et år holdt kløften op med at vokse. Den sårbare gruppe var allerede skudt ned, men den sakkede ikke længere bagud i forhold til resten af sektoren.

Så kom det andet brud.

I løbet af 2024 og 2025, mens udbredelsen af generativ AI tog fart, faldt en million dollars i medianvirksomheden med høj eksponering til omkring 400.000. Den tilsvarende med lav eksponering nåede omkring 1,08 millioner.

Den rækkefølge er langt mere interessant end blot at sige, at AI slog edtech ihjel. Først faldt de mest udsatte forretninger kraftigt, længe før ChatGPT. Så stabiliserede de sig. Og da folk fik en markant lettere vej til mange af de samme resultater, åbnede kløften sig igen.

Det beviser ikke, at AI stod bag hver eneste dollar af det andet fald. Aktiekurser afhænger af renter, eksekvering, konkurrence, geografi og udgangspunktet for værdiansættelsen. Men det fastlægger rækkefølgen af begivenhederne, og den rækkefølge gør det svært at holde fast i den bekvemme forklaring.

Havde generativ AI knækket en ellers sund branche fra den ene dag til den anden, ville det første fald være kommet efter generativ AI. Det gjorde det ikke.

Den mere ubehagelige konklusion er, at en stor del af edtech allerede havde bygget en skrøbelig økonomi omkring produkter, kunderne var svære at fastholde på, som blev brugt for lidt, eller som afhang af, at folk igen og igen vendte tilbage til separate digitale destinationer. Markedets første korrektion afslørede de svagheder. I en kort periode stabiliserede skaden sig. Så ændrede generativ AI adgangen til mange af de samme resultater, og de virksomheder, hvis værditilbud var lettest at kopiere, faldt igen.

Den forskel betyder meget. Var ChatGPT det oprindelige problem, ville løsningen bare være at putte AI i det eksisterende produkt. Men hvis behovet består, mens vejen dertil ændrer sig, bliver det strategiske spørgsmål meget større: skal den samme værdi overhovedet fortsat leveres gennem det samme produkt?

Hvad hvis produktet er rigtigt, men indpakningen forkert?

Netflix vandt ikke, fordi folk pludselig ville se andre film, men fordi ny teknologi gjorde det muligt at ændre selve leveringen og fjerne friktion for brugeren. I dag er den friktion sjældent en fysisk barriere. Det kan være en betalingsmodel, endnu et login, en app man skal huske at åbne, et kursus man skal forpligte sig til, eller en platform der kræver, at man ændrer en vane, man allerede har.

Læringsvirksomheder konkurrerer i dag med den læring, der foregår på TikTok, i podcasts, på YouTube og i generel AI. I et europæisk studie fra 2025, bestilt af YouTube, svarede 74 % af de unge adspurgte, at de havde set YouTube-videoer for at lære noget nyt til skolen. Sesame Streets officielle YouTube-kanaler fik over fem milliarder visninger i året op til, at YouTube annoncerede et udvidet samarbejde, der fra 2026 gør platformen til det største digitale hjem for programmets afsnit.

Igen: efterspørgslen efter læring forsvandt ikke. Den flyttede sig. Og vælger man den forkerte indpakning, kan det ende helt galt. Et nyere dansk eksempel gør det usædvanligt konkret. MYiNNERME var en digital app til mental sundhed, udviklet af børnepsykologer til børn, unge og deres familier. Den tilbød selvhjælpsforløb om angst, vrede, generthed, sorg, mobning, søvn og selvværd og oversatte veletablerede psykologiske metoder til øvelser og indhold, familier kunne bruge uden om et traditionelt terapiforløb. Firmaet bag MYiNNERME er lukket og har meldt ud, at de aldrig fandt en rentabel forretningsmodel.

Den oplagte konklusion er, at produktet fejlede, fordi der ikke var nok efterspørgsel efter det. Jeg er ikke sikker på, at det er den rigtige konklusion.

Stod jeg med samme problem i dag, ville jeg spørge, om en app overhovedet skulle være udgangspunktet. Forestil dig i stedet en ansigtsløs YouTube-kanal bygget op omkring virksomhedens figurer. Et barn kunne møde en af figurerne i en video om angst, genkende den i en historie om mobning, bruge den til en vejrtrækningsøvelse og senere støde på den igen i indhold om at falde i søvn eller starte i ny skole.

At bygge det univers kræver slet ikke den produktionsinfrastruktur, det ville have krævet for få år siden. Figurer kan skabes og animeres med AI-værktøjer. Manuskripter, stemmer, oversættelser og varianter kan laves langt billigere. Produktion og udgivelse kan automatiseres. De samme figurer og psykologiske begreber kan vandre fra lange YouTube-videoer til shorts, til lyd, til printbare øvelser og til materiale, psykologer eller skoler kan bruge.

De kan også forlade skærmen helt. Den samme rettighed kunne blive til historiebøger, malebøger trykt efter behov, aktivitetshæfter, kort eller andre fysiske produkter, uden at virksomheden skal binde kapital i store lagre, før den ved, om nogen overhovedet vil have dem.

Pointen er ikke, at MYiNNERME nødvendigvis ville have klaret sig som YouTube-virksomhed. Pointen er, at det værdifulde aktiv aldrig var appen. Det var den kliniske erfaring, de psykologiske rammer, øvelserne, historierne, figurerne og evnen til at oversætte svære følelser til noget, børn kan forstå. Appen var bare én indpakning af den værdi. I dag findes der mange flere.

En konkret virksomhed og en konkret leveringsmodel kan altså forsvinde, mens den underliggende efterspørgsel består. Faktisk kan den efterspørgsel i dag mødes flere steder end nogensinde: i dedikerede produkter, hos content creators, på YouTube, via AI, i skolerne, mellem terapisessioner og i fysiske produkter koblet til digitale universer.

Men det skaber et nyt problem. Bliver software billigere at bygge og indhold billigere at skabe, er godt indhold heller ikke nok i sig selv.

Det afgørende spørgsmål bliver: hvad er stadig knapt? Det kan være et troværdigt brand. Det kan være eksamensbeviser og certificeringer. Det kan være egne data, distribution, institutionelle relationer, fællesskab, regulatorisk godkendelse, integration i en arbejdsgang, anerkendt prøvning, rettigheder eller dokumenterede resultater. Svaret varierer fra virksomhed til virksomhed, men princippet er det samme. Når software og indhold bliver billigere at producere, må investoren finde den del af forretningen, som ikke er let at kopiere. Det er et langt mere brugbart spørgsmål end at spørge, om en edtech-virksomhed “bruger AI”.

Edtechs dvd-øjeblik er også et problem for investorerne

Generativ AI har ikke bare skabt nye steder at lære. Den har også gjort det langt billigere for virksomheder at pakke den samme viden ind til alle de nye steder.

Men der er en vigtig forskel på MYiNNERME, edtech-branchen bredt og Netflix. Edtech går ikke ind i denne teknologiske omvæltning fra en stærk økonomisk position. Problemerne og værdifaldene lå der allerede, længe før generativ AI. Længe før ChatGPT var et seriøst alternativ, sad sektoren med ubrugte licenser, svag evidens for effekt, dårlig fastholdelse og forretningsmodeller, der afhang af, at folk igen og igen kom tilbage til separate digitale destinationer. De børsnoterede markeder havde allerede indregnet en del af de svagheder.

Derfor tror jeg i stigende grad, at den vigtige skelnen ikke bare går mellem “AI-virksomheder” og “ikke-AI-virksomheder”. Den går snarere mellem destinationer og infrastruktur. En destination kræver, at kunden aktivt beslutter sig for at komme til dig. Åbne appen. Logge på platformen. Starte kurset. Se lektionen. Spørge tutoren. Gennemføre aktiviteten. Komme igen i morgen.

Infrastruktur er noget andet. Den lever inde i en arbejdsgang, en institution eller et system, der allerede findes. Eksamensbeviser, prøvning, udgivelse, compliance, skoleadministration, anerkendte grader, egne data og dybt integrerede institutionelle værktøjer kan have en form for værdi, som en generel AI-grænseflade har sværere ved at snuppe.

Det gør ikke infrastrukturforretninger usårlige, og det gør ikke destinationsforretninger værdiløse. Men skelnen betyder noget, fordi AI dramatisk sænker friktionen ved at få information, forklaringer, træning og stadig mere personlig undervisning, uden at man behøver besøge en specialiseret platform.

For en investor er spørgsmålet derfor ikke bare, om folk bliver ved med at lære. Det gør de indlysende nok. Spørgsmålet er, hvem der opsnapper økonomien i den læring, når selve grænsefladen ændrer sig.

Sammenligningen med det børsnoterede marked gør problemet svært at overse. At følge markedets oprindelige hierarki blandt de største børsnoterede uddannelsesvirksomheder uden Kina gjorde en million dollars til omkring 1,02 millioner. I samme periode blev den samme million i SPY til omkring 2,10 millioner. Specialistinvestoren skulle forstå pædagogik, offentlige indkøb, implementering, regulering og en række meget specifikke forretningsmodeller, mens den passive investor bare kunne købe en indeksfond og gå til frokost. Den passive løsning endte med mere end dobbelt så mange penge.

Det er ikke et argument for, at uddannelse er værdiløst. Nærmest tværtimod. Folk lærer stadig. Børn har stadig brug for hjælp med angst. Voksne skal stadig forstå investering. Medarbejdere skal stadig have nye kompetencer. Studerende har stadig brug for forklaringer. Efterspørgslen forsvinder ikke. Det ubehagelige spørgsmål er, hvorfor så mange af de virksomheder, der blev bygget til at møde den efterspørgsel, har været så dårlige til at omsætte den til værdi for aktionærerne.

MYiNNERME-sagen gør det ekstra tydeligt. Virksomheden kan lukke, mens den viden, den skabte, de problemer, den løste, og menneskerne, der stadig søger løsninger, alle sammen består. Den værdi kan i dag pakkes ind i et YouTube-univers, tilbagevendende figurer, samtaler med AI, bøger, skoleforløb, værktøjer til psykologer, fysiske produkter eller enhver kombination, ofte til en brøkdel af den produktions- og distributionsomkostning, det ville have kostet for få år siden.

Filmen mistede ikke sin værdi, da dvd'en gjorde. Men det var heller ikke meget værd at eje en dvd-forretning, efter streaming kom til.

Det er den skelnen, jeg mener betyder noget for edtech-investorer i dag. Godt indhold er ikke nok. God pædagogik er ikke nok. Selv massiv efterspørgsel er ikke nok. Ændrer teknologien måden, folk konsumerer den underliggende værdi på, må virksomheden flytte sig med den og finde ud af, hvordan den opsnapper nok værdi til at retfærdiggøre den investerede kapital.

Og hvis den ikke kan det, hvorfor så løbe den ekstra risiko? Kan man lægge en million dollars i et passivt indeks, stort set ikke løfte en finger og ende med det dobbelte, må specialiseret edtech-kapital gøre sig fortjent til at eksistere.

Derfor bliver den næste generations vindere ikke bare dem, der bruger AI, har de bedste kurser eller gør nutidens apps en tand bedre. Det bliver dem, der forstår, hvilken del af deres produkt der stadig er reelt knap og værdifuld, hvordan den nye teknologi ændrer den billigste og mest naturlige måde at levere den på, og hvor i det nye system der overhovedet ligger en forretning, som kan opsnappe nok værdi til at give et attraktivt afkast.

Derfor tror jeg, at 90 % af edtech kan forsvinde, uden at læringen forsvinder med den. Vi løber ikke tør for ting at lære. Vi løber måske bare tør for grunde til at betale for dem på de måder, edtech har brugt det seneste årti på at bygge.

Serie · Sådan forsvinder 90 % af edtech

En serie i fire dele om pengene, markedet og modellerne, der gjorde en hel branche til noget andet.

Del 1 af 4Hvis jeg drev en edtech-fond, ville jeg skide i bukserne

Del 2 af 4 · Forrige i serienDin edtech-investor vil have et åbent forhold

Del 3 af 4Edtech sender stadig dvd'er med posten

Del 4 af 4 · Næste i serienNår din kunde bliver din konkurrent: Kunden fandt en genvej til softwarefabrikken

Læs hele serien →Beslægtede essays

AI & edtech · 18. juli 2026Hvis jeg ser én til pushe “PedTech”, kaster jeg opPædagogikken har aldrig manglet som idé. Det svære har været at bygge en branche, hvor økonomien lod pædagogikken blive stående i centrum, når investorerne bankede på.

AI & samfund · 8. september 2026Elon Musk købte den forkerte ejendom til alt-appenMusk købte Twitter for at bygge alt-appen. Men mens han samlede opmærksomhed, samlede OpenAI intentioner, og det kan være den mest værdifulde adresse på internettet.

AI & samfund · 16. august 2026AI åd internettet. Nu skal den afgøre, hvem der er menneskeVandmærkning skal gøre syntetisk indhold synligt. Den kan også gøre AI-selskaber til dem, vi beder om at attestere menneskeligt forfatterskab.

← Tilbage til bloggenLæs første delOplægInviter mig til at holde oplæg

---

### https://sahra-josephine.com/da/blog/edtech-open-relationship

← Blog

AI & Edtech·30. juli 2026Din edtech-investor vil have et åbent forhold

Historien om, hvordan 90 % af edtech forsvinder. Du er stadig i porteføljen, du er bare ikke længere deres eneste type.

Denne tekst er maskinoversat fra engelsk. Nuancer kan gå tabt undervejs, så vi anbefaler at læse originalen. Læs den engelske original

Det mærkeligste ved langsomt at blive droppet er, at ingen egentlig slår op med dig. De svarer stadig på dine beskeder. Dit billede står måske stadig fremme. Deres tandbørste ligger måske stadig på dit badeværelse. Men sproget skifter. De taler ikke helt om fremtiden på samme måde, beskrivelsen af, hvad de leder efter, bliver bredere, og pludselig opdager du, at de har skiftet type uden at fortælle dig det.

Sådan noget foregår lige nu mellem edtech og branchens investorer.

I min forrige tekst spåede jeg, at mellem 90 og 95 procent af nutidens edtech-virksomheder vil lukke, blive konsolideret eller ende reelt irrelevante. Den spådom hviler på flere forskellige forandringer, som ikke kan rummes ordentligt i én tekst, så jeg tager dem én ad gangen.

Jeg havde regnet med at starte med styrtdykkende værdiansættelser eller kunstig intelligens, der overtager uddannelsesprodukter. I stedet fandt jeg noget mere stilfærdigt og på sin vis mere afslørende. Edtech er ikke nødvendigvis blevet droppet. Investoren vil bare have et åbent forhold.

Pengene forsvandt først

Den globale venturekapital i edtech toppede med 20,8 milliarder dollar i 2021. I 2024 var tallet faldet til 2,4 milliarder dollar, et fald på omkring 89 procent. Kun en svag bedring bragte tallet op på 2,6 milliarder dollar i 2025.

Kunstig intelligens har ikke alene skabet det kollaps. Faldet begyndte, før generativ AI slog igennem, og meget af det var en uundgåelig korrektion, efter at pandemiefterspørgsel, billige penge og vanvittige værdiansættelser pumpede et boom op, der aldrig kunne holde.

Men fire år efter toppen ligger de klassiske edtech-investeringer stadig tæt på det laveste niveau i et årti. Kapitalen er ikke bare kommet tilbage på de gamle vilkår. Den er blevet mere kræsen, mere knyttet til beskæftigelse og langt mere interesseret i produkter, der kan kaldes infrastruktur, produktivitet eller kunstig intelligens. Og så begyndte navnene at skifte.

Edtech skiftede datingprofil

Jeg gennemgik den offentlige positionering hos ti investorer, der historisk er knyttet til uddannelse og edtech. Jeg valgte dem netop på grund af den forbindelse, ikke fordi deres nuværende sprogbrug passede godt ind i min pointe. De er ikke et statistisk repræsentativt udsnit af hele venturemarkedet, og en hjemmeside afslører ikke enhver beslutning i en investeringskomité. Men hjemmesider viser, hvordan fonde ønsker at blive forstået af stiftere, porteføljeselskaber og deres egne investorer.

Mønsteret var ikke universelt. Owl Ventures, GSV Ventures og Rethink Education står stadig tydeligt forankret i uddannelse. Educapital kalder sig fortsat en edtech- og future-of-work-fond. Det er værd at nævne, for jeg ville ikke bare pille de investorer ud, der havde lagt afstand til ordet edtech.

Men flere af de mest kendte specialiserede investorer taler nu om et langt bredere forhold. Reach Capital startede i 2015 med at investere i krydsfeltet mellem teknologi og uddannelse. Fondens fornyede tese dækker nu læring, sundhed og arbejde under ét.

Brighteye kalder sig ikke længere først og fremmest en edtech-investor. Fonden bakker op om stiftere, der bygger det, den kalder “HumanOS”, altså systemer der hjælper mennesker med at lære, arbejde og løbende tilpasse sig. Selv flagskibsrapporten hedder ikke længere en edtech-fundingrapport. Den hedder nu Learning & Work Funding Report.

Emerge blev grundlagt som Europas eneste specialiserede edtech-fond. I dag kalder fonden sig en fond for fremtidens arbejde og læring, og den investerer i alt fra dagtilbud til karrierevejledning og AI på arbejdspladsen.

Kaizenvest skriver, at det, der begyndte som Indiens første uddannelsesfokuserede kapitalfond, i dag er en “omfattende strategi for menneskelig økonomisk mobilitet”. Fondens nuværende tese spænder over uddannelse, sundhed, finansiel inklusion og jobskabelse.

Learn Capital taler stadig meget om uddannelse, men kalder nu sit felt for uddannelse og udvikling af human capital, herunder AI-drevet læring, opkvalificering, karriereudvikling og trivsel.

Ordvalget varierer, men retningen er slående ens. Edtech bliver til læring. Læring bliver til kompetencer. Kompetencer bliver til arbejdsstyrkeudvikling. Arbejdsstyrkeudvikling bliver til human capital, karrieremobilitet, produktivitet eller menneskeligt potentiale. Uddannelse er der stadig, den er bare ikke længere alene om opmærksomheden.

Det er strategisk fornuftigt nok. En dedikeret edtech-fond skal finde attraktive investeringer inden for edtech. En bredere tese kan følge den samme lærende videre ind i beskæftigelse, sundhed, rekruttering, produktivitet eller AI. Fonden beholder sin ekspertise, men får samtidig et langt større felt af virksomheder og budgetter at gå efter.

Set fra investorens stol er det bare diversificering. Set fra stiftersædet betyder det, at den næste check ikke længere behøver gå til en som dig.

Så tjekkede jeg min egen investor

Først efter at have set det brede mønster kiggede jeg ordentligt på Sparkmind, en af CanopyLABs investorer. Jeg betragter ikke branchen udefra. Jeg har brugt over ti år på at bygge CanopyLAB, som ligger midt i den del af edtech, der er mest udsat for AI: platforme, der skaber, organiserer og leverer læringsindhold.

I 2020 blev Sparkmind beskrevet som en nordisk venturefond med speciale i edtech. Planen var at investere hele vejen gennem uddannelsesrejsen, fra dagtilbud til livslang læring og virksomhedstræning.

Besøg Sparkminds hjemmeside i dag, og du møder to døre: Human Capital og Security. CanopyLAB og resten af den oprindelige uddannelsesportefølje ligger under Human Capital. Den del af fonden har investeret i 26 virksomheder og laver ikke længere nye førstegangsinvesteringer, hvilket måske bare er en almindelig venturefonds livscyklus. Sparkmind siger, at man fortsat vil bakke sine porteføljeselskaber op.

Set fra Sparkminds stol er logikken klar. Uddannelse, beskæftigelse, AI og geopolitisk sikkerhed forandrer sig samtidig, og et bredere mandat giver flere steder at investere. Men strategisk logik gør ikke konsekvensen mindre til at mærke for stifteren. Fornemmelsen er umiskendelig. Jeg er stadig i porteføljen, men edtech er ikke længere det ord, der forklarer selskabets fremtid. Forholdet fortsætter, investoren har bare skiftet type.

Bedringen dukker op, når kategorien bliver bredere

Det nye sprog ville være mindre interessant, hvis det bare var branding. Fundingtallene peger på noget mere håndgribeligt. I Europa mere end fordobledes investeringerne i den bredere kategori “Learning & Work”, fra 710 millioner euro i 2024 til 1,6 milliarder euro i 2025. Det lyder som en voldsom bedring.

Men Brighteyes egen opdeling viser, at klassisk edtech, altså skoler, videregående uddannelse og individuel livslang læring, kun fik 471 millioner euro. Virksomheds- og arbejdspladslæring fik 601 millioner euro, og resten af den udvidede kategori dækker produktivitets-, rekrutterings- og talentplatforme, som sjældent selv beskriver sig som læringsprodukter.

Det gør ikke det store tal forkert. Læring bliver rent faktisk vævet sammen med arbejde og performance. Men det ændrer, hvad bedringen egentlig betyder.

Edtech-finansieringen er stadig hårdt presset. Learning & Work-finansieringen kommer sig. Bedringen ser meget mere imponerende ud, når man udvider kategorien til at omfatte virksomheder, som edtech-investorer næppe ville have kaldt edtech for fem år siden. Måske vender markedet ikke tilbage til edtech. Måske flytter edtech bare derhen, hvor markedet er taget hen.

Så gør edtech sig selv om for at forblive attraktiv

Også de børsnoterede selskaber skifter sprog. Guild Education blev til Guild og satte karrieremobilitet centralt i sin fortælling. Chegg kalder nu sin akademiske forretning for “legacy Academic Services”, fremhæver Skilling som sin vækstmotor og er gået ind i træning af AI-modeller. Coursera kaldte sammenlægningen med Udemy for en omfattende kompetenceplatform til AI-æraen, mens Multiverse i dag kalder sig Europas AI-adoptionsplatform. Produkterne underviser måske stadig folk, men salgstalen handler i stigende grad om karrieremobilitet, produktivitet, performance på arbejdsstyrken eller AI-adoption. Det er vidt forskellige beslutninger, og de beviser ikke, at nogen af virksomhederne får succes. Men tilsammen viser de, at uddannelse fylder mindre og mindre i det sprog, virksomhederne bruger til at tiltrække kunder og kapital.

Hos børsnoterede selskaber er det relativt let at følge et kategoriskifte. Investorpræsentationer, regnskabsopkald, opkøb og skiftende forretningsområder ligger åbne. Private virksomheder er sværere at læse. Deres værdiansættelser bliver ikke løbende testet af markedet, og de strategiske samtaler mellem stiftere, bestyrelser og investorer forbliver bag lukkede døre. Derfor er deres hjemmesider ekstra afslørende. De viser, hvilken del af virksomheden der er valgt ud til udstillingsvinduet.

Efter Sparkmind blev jeg nysgerrig på de andre selskaber i porteføljen ved siden af CanopyLAB. To af dem viser meget forskellige udviklingsspor. Ingen af dem beviser, at virksomheden overlever. Begge viser, hvordan private edtech-selskaber allerede ændrer deres offentlige identitet og deres plads i teknologistakken.

Female Invest vælger ikke længere edtech-kategorien

Sparkmind kategoriserer Female Invest som en virksomhed inden for livslang læring. I 2021 blev selskabet beskrevet som en “EdTech-platform og et fællesskab”, der brugte abonnementslæring til at hjælpe kvinder med at forstå privatøkonomi og investering.

Produktet er stadig soleklart uddannelse. Medlemmerne får kurser, korte lektioner, finansnyheder, budgetværktøjer, adgang til eksperter og en virtuel handelssimulator. Det, der har ændret sig, er den kategori, Female Invest vælger at vise frem for markedet.

Besøg Female Invests hjemmeside i dag, og overskriften handler hverken om edtech, digital læring eller gennemførte kurser. Female Invest kalder sig “the money app for every step of your journey”. Budskabet til kvinder er, at de skal tage kontrol over egen økonomi, øve sig i at investere og opbygge formue. Fællesskabet, som ifølge selskabet tæller over 85.000 kvinder fra 125 lande, sælges som en del af en bevægelse, der skal lukke det økonomiske kønsgab.

Forskellen er ikke ligegyldig. Female Invest sælger ikke bare en bedre måde at lære om investering på. Selskabet sælger selvtillid, økonomisk uafhængighed, adgang til eksperter, identitet og medlemskab af et fællesskab bygget op om et uløst samfundsproblem.

Kategorien har flyttet sig i årevis. I 2022 opkøbte Female Invest Gaia Investments, en platform for bæredygtige investeringer, med planer om at bygge egentlig handel ind i produktet. Det nuværende tilbud ser dog mest ud til at fokusere på virtuel handel og undervisning frem for at fungere som en rigtig børsmægler, så jeg vil ikke kalde selskabet en investeringsplatform i dag. Men opkøbet afslørede den strategiske ambition om at bevæge sig fra at lære folk om penge til at hjælpe dem handle på det, de har lært.

I 2023 forklarede Female Invest, at selskabets nye brandidentitet skulle række ud over økonomien og blive en tungere stemme i debatten om ulighed mellem kønnene i det hele taget. I 2024 annoncerede selskabet en Series A på 11 millioner dollar. Nogle medier kaldte det en edtech-virksomhed. Andre kaldte det fintech. Den nuværende branding får begge etiketter til at virke for snævre.

Det virker bevidst. Female Invest har ikke brug for, at kunderne skal beslutte, om selskabet hører hjemme i fintech, edtech, medier eller fællesskab. Brandet er bygget op om det problem, det vil løse, ikke om den softwarekategori, det bruger til at løse det.

Female Invest består af praktiske værktøjer, der bringer brugeren tættere på handling. Simulatoren gør, at man kan øve sig, før man tager en reel økonomisk risiko. Fællesskabet giver social opbakning. Eksperterne skaber tillid på et felt, hvor fejl koster. Sammen kan de elementer skabe en relation, der er sværere at erstatte end et almindeligt katalog af finanskurser.

Men intet af det beviser, at Female Invest overlever. Undervisningsindhold, markedsnyheder og introduktionsguides er alle stærkt udsatte for AI. Et fællesskab holder kun, hvis folk rent faktisk deltager, værdsætter relationerne og ikke nemt kan genskabe det samme et andet sted. Female Invest er derfor ikke bevis på, at formål og fællesskab garanterer overlevelse. Det er bevis på, at en privat edtech-virksomhed kan gøre uddannelse til blot én del af en langt større identitet. Virksomheden kan blive ved med at vokse, mens den kategori, investoren bruger til at beskrive den, bliver mere og mere ligegyldig for dem, der rent faktisk køber produktet.

imagi er ved at blive laget mellem skoler og AI

Det andet eksempel er imagi, tidligere kendt som imagiLabs. I 2022 beskrev imagiLabs sig selv som en edtech-startup, der lancerede en platform, hvor undervisere kunne undervise i Python. Økosystemet bestod af en gamificeret mobilapp, en wearable ved navn imagiCharm og et pensum bygget op om, hvad piger i præteenagealderen interesserer sig for. Idéen var enkel nok: imagi havde bygget sit eget miljø til at lære børn at kode. Det er ikke længere kernen i historien.

I juli 2026 rejste imagi 4,5 millioner dollar til at bygge det, virksomheden kalder “det sikre uddannelseslag mellem AI-værktøjer og skoler”. I stedet for at bede eleverne om at lære udelukkende inde i et imagi-produkt giver virksomheden dem nu superviseret adgang til generelle AI-værktøjer via pensum, lærerstøtte og sikkerhedskontroller.

Partnerne er Lovable og OpenAI. Gennem samarbejdet mellem Lovable og imagi bruger eleverne AI til at bygge applikationer, mens imagi leverer klasseværelsesadgang, undervisningsplaner, automatisk oprettelse af elever og læreruddannelse. Virksomheden siger, at flere frontier-AI-værktøjer er på vej ind.

Det er et næsten bogstaveligt eksempel på den forandring, jeg beskrev i min forrige tekst. Læringsoplevelsen flytter ind i det generelle AI-miljø. Edtech-virksomheden gør sig selv til det pædagogiske, administrative og sikkerhedsmæssige lag mellem det miljø og den lærende.

Ændringen er ikke bare kosmetisk. imagi er gået fra at ville eje hele destinationen til at administrere adgang til teknologier, som langt større virksomheder ejer. Selskabet underviser ikke længere kun i Python gennem sin egen brugerflade. Det hjælper skoler med at beslutte, hvordan børn trygt kan lære med værktøjer, de måske allerede bruger andre steder.

De tidlige tal ser lovende ud, om end de kommer fra virksomheden selv. imagi siger, at man er til stede i over 100 skoledistrikter, har nået mere end 700.000 elever i 140 lande, har tredoblet den årlige tilbagevendende omsætning og tredivedoblet antallet af brugere på et år. Den nye runde talte Sparkmind sammen med Morgan Stanley og private investorer fra blandt andre ElevenLabs, GitHub, Spotify og Lovable.

Igen: det beviser ikke, at imagi overlever. Men det viser, hvorfor den nye position kan være attraktiv.

Skoler har brug for mere end adgang til en model. De skal bruge børnesikkerhed, privatliv, compliance, klasserumsledelse, læreruddannelse, pensum og nogen, der tager ansvaret, når noget går galt. Frontier-AI-selskaberne har måske ikke lyst til at skræddersy alt det til hvert eneste skoledistrikt, hver aldersgruppe og hvert regelsæt. imagi forsøger at fylde det hul ud.

Hvis virksomheden kan eje de betroede skolerelationer, pædagogikken, implementeringen og evnen til at koble flere AI-leverandører sammen, kan den blive værdifuld infrastruktur. Den behøver ikke bygge den stærkeste model. Den skal bare være den sikreste og mest brugbare vej for skoler til at få adgang til de modeller, der ender med at betyde noget.

Men afhængigheden er tydelig. OpenAI, Google, Anthropic, Microsoft, Lovable eller en af de store skoleplatforme kunne sagtens bygge det meste af det lag selv. De har større distribution, mere kapital og kontrol over selve teknologien. Jeg tvivler også på, at markedet har brug for hundredvis af uafhængige virksomheder, der sidder mellem skoler og frontier-AI. Nogle få kan bygge forsvarlige positioner omkring geografi, aldersgrupper, regulering eller specifikke pædagogiske behov. Mange andre risikerer at ende som udskiftelige skaller udenpå produkter, de ikke selv styrer.

Det er derfor, imagi er et så interessant eksempel. Udviklingen kan være strategisk klog og fuldt ud på linje med den oprindelige mission om at hjælpe børn, især piger, med at blive skabere af teknologi frem for bare forbrugere af den. Men virksomhedens arkitektur har ændret sig. Den afhænger nu af det generelle AI-økosystem, den engang måske forventede at konkurrere med.

Selskabet har ikke opgivet uddannelse. Det har accepteret, at uddannelse i stigende grad kan foregå et andet sted, og forsøger i stedet at blive det lag, der gør det muligt.

Female Invest og imagi fortæller os ikke, hvilke virksomheder der overlever. De viser os, hvordan private virksomheder allerede reagerer på de samme kræfter, som er tydelige hos de børsnoterede selskaber og investorerne. Den ene har gjort kategorien underordnet formål, identitet og fællesskab. Den anden er gået fra at være selve læringsdestinationen til at blive infrastruktur omkring en andens AI-platform. Begge kan vise sig at være glimrende beslutninger. Begge kan slå fejl. Beviset ligger ikke i resultatet. Beviset ligger i bevægelsen.

En branche kan forsvinde, uden at alle dør

Jeg havde oprindeligt forestillet mig, at edtech skulle forsvinde gennem en række synlige nederlag: virksomheder der lukker, værdiansættelser der styrtdykker, platforme der bliver forældede. Alt det kommer til at ske. Men brancher kan også forsvinde ved simpelthen at blive opslugt.

Virksomheden overlever, men bliver en arbejdsmarkedsvirksomhed. Produktet overlever, men læring bliver bare én funktion inde i en AI- eller produktivitetsplatform. Investoren overlever, men uddannelse bliver ét mulige udtryk blandt mange for en langt bredere investeringstese.

Venturekapital har aldrig været et ægteskab. Det handlede aldrig om i medgang og modgang. Investoren ville altid følge det bedste afkast og til sidst finde en exit. Stifteren var måske gået ind af en helt anden grund.

De færreste bruger år på at bygge i uddannelse for at maksimere en kategorimultipel. De starter, fordi de er optaget af elever, lærere, pædagogik, adgang eller menneskeligt potentiale. De vil inspirere folk, åbne muligheder og hjælpe nogen med at se, hvad de kan blive til. Men når nok ekstern kapital er kommet ind, får virksomheden endnu et formål: at beskytte aktionærværdien.

Når edtech holder op med at tiltrække kapital, kan det pres skubbe en virksomhed væk fra sin kerne. Uddannelse bliver til kompetencer. Læring bliver til produktivitet. Elever bliver til human capital. Den oprindelige mission bliver stående på om os-siden, mens det kommercielle tyngdepunkt flytter derhen, hvor pengene stadig er.

Nogle gange er det legitim tilpasning. Nogle gange er det den eneste vej til at overleve. Men product-market fit bør fortælle en virksomhed, hvordan den skal indfri sit formål, ikke hvad formålet er.

Venturekapitalen går videre. Edtech-virksomheden bliver tilbage og laver sig om til noget, markedet måske vil elske næste gang. Sådan forsvinder edtech: ikke ved at alle virksomheder dør, men ved at overlevelse kræver, man glemmer, hvorfor man blev født.

Serie · Sådan forsvinder 90 % af edtech

En serie i fire dele om pengene, markedet og modellerne, der gjorde en hel branche til noget andet.

Del 1 af 4 · Forrige i serienHvis jeg drev en edtech-fond, ville jeg skide i bukserne

Del 2 af 4Din edtech-investor vil have et åbent forhold

Del 3 af 4 · Næste i serienEdtech sender stadig dvd'er med posten

Del 4 af 4Når din kunde bliver din konkurrent: Kunden fandt en genvej til softwarefabrikken

Læs hele serien →Beslægtede essays

AI & edtech · 18. juli 2026Hvis jeg ser én til pushe “PedTech”, kaster jeg opPædagogikken har aldrig manglet som idé. Det svære har været at bygge en branche, hvor økonomien lod pædagogikken blive stående i centrum, når investorerne bankede på.

AI & samfund · 8. september 2026Elon Musk købte den forkerte ejendom til alt-appenMusk købte Twitter for at bygge alt-appen. Men mens han samlede opmærksomhed, samlede OpenAI intentioner, og det kan være den mest værdifulde adresse på internettet.

AI & samfund · 16. august 2026AI åd internettet. Nu skal den afgøre, hvem der er menneskeVandmærkning skal gøre syntetisk indhold synligt. Den kan også gøre AI-selskaber til dem, vi beder om at attestere menneskeligt forfatterskab.

← Tilbage til blogLæs den første tekstOplægBook mig til et oplæg

---

### https://sahra-josephine.com/da/blog/edtech-fund

← Blog

AI & Edtech·25. juli 2026Hvis jeg drev en edtech-fond, ville jeg skide i bukserne

Denne tekst er maskinoversat fra engelsk. Nuancer kan gå tabt undervejs, så vi anbefaler at læse originalen. Læs den engelske original

Jeg har bygget i edtech i over ti år.

Jeg var med til at stifte CanopyLAB, en venturefinansieret læringsteknologivirksomhed, fordi jeg troede på, at adaptiv læring for alvor ville forandre uddannelse. Det tror jeg faktisk stadig, jeg havde ret i.

Det, jeg undervurderede, var, hvor lang tid teknologien skulle bruge på at indhente de pædagogiske principper.

I mange år var adaptiv læring en bedre idé end brugeroplevelse. Vi kunne sagtens indsamle data, anbefale indhold og designe forskellige veje gennem et kursus. Men de veje skulle bygges og vedligeholdes, og det krævede alt for meget manuelt arbejde fra underviserne. Teknologien kunne understøtte tilpasning, men den forstod ikke rigtig den lærende, kunne ikke generere det rette materiale og kunne ikke løbende omforme oplevelsen.

Den var adaptiv, ja, men kun inden for rammer, andre allerede havde sat.

Så kom kunstig intelligens, og det begyndte at ændre sig.

I 2019 opfandt jeg AICATO, som CanopyLAB lancerede som verdens første AI-værktøj til kursusudvikling. Det automatiserede dele af kursusproduktionen, der før krævede timevis af menneskeligt arbejde. Et tidligt forsøg på at løse et af adaptiv lærings grundproblemer: du kan ikke skabe en reelt individuel læringsoplevelse, hvis alt indhold og alle mulige veje først skal produceres i hånden.

Så kom generativ AI, og pludselig kunne teknologien meget mere end bare vælge mellem foruddefinerede muligheder. Den kunne generere, forklare, spørge, justere og svare i realtid. Endelig havde teknologien indhentet idéen.

Det burde gøre mig usædvanligt optimistisk på edtechs vegne. I stedet gør det mig nervøs.

I dag giver jeg virksomheder det modsatte råd

For ti år siden ville jeg næsten altid have svaret nej, hvis en virksomhed spurgte, om den skulle bygge sin egen læringsplatform. Software var dyrt og tog tid at bygge. Man skulle bruge specialiserede udviklere, designere, produktchefer, infrastruktur og løbende vedligehold, og selv hvis det lykkedes, blev resultatet sjældent nær så godt som produktet fra en virksomhed, der ikke lavede andet.

Dengang var det fornuftige råd at finde en specialiseret leverandør, sende opgaven i udbud, købe det bedste produkt på markedet og tilpasse det efter behov. I dag giver jeg oftere det modsatte råd, og det gælder langtfra kun edtech. Jeg har selv knap 100.000 abonnenter fordelt på mine engelske og spanske mailinglister. Bare det at vedligeholde og sende mails til de 36.000 mennesker på den spanske liste kostede mig omkring 10.000 kroner om måneden (1.520 dollar) gennem Mailchimp. Det er mange penge for at maile til en liste, jeg i forvejen ejer. Så i sidste uge byggede jeg mit eget afsendelsesværktøj i Lovable og koblede det til SendGrid. Det tog under 30 dollar at bygge og koster nu 25-30 dollar om måneden at drive. Hvorfor skulle jeg blive ved med at betale for et softwareabonnement?

Nu skal ikke alle virksomheder bygge alle deres værktøjer selv. Små virksomheder vinder stadig ved at købe noget færdigt, og komplekse, regulerede systemer er en helt anden snak. Men har man 30.000, 40.000 eller 100.000 brugere, ser regnestykket pludselig markant anderledes ud.

Et værktøj, der før krævede en softwarevirksomhed, et udviklerteam og venturekapital, kan i dag ofte laves af én person på en weekend. Nogle virksomheder tjener udviklingsomkostningen hjem på en til to måneders sparede abonnementer. Det udfordrer en af de bærende antagelser bag hele software-as-a-service-økonomien. Vi brugte år på at bevæge os fra eje til leje, fordi specialiseret software var billigere og bedre, end de fleste organisationer selv kunne bygge. Nu falder prisen på at bygge, mens prisen på at leje står stille.

Edtech har et andet, endnu større problem

Edtech er ikke kun sårbar, fordi organisationer selv kan bygge mere software. Branchen er lige så sårbar, fordi de lærende måske slet ikke længere har brug for en separat læringsplatform. Elon Musk har talt om at gøre X til en everything app i årevis. Grundidéen tror jeg var rigtig nok, men udgangspunktet var forkert.

Everything-appen var nok aldrig et socialt netværk. Den bliver den AI-platform, der allerede ved, hvad du arbejder med, hvad du interesserer dig for, hvad du forstår, og hvor du gang på gang går i stå. Om det platform hedder ChatGPT, Claude, Gemini, Grok eller noget helt fjerde, der endnu ikke er udgivet, er ligegyldigt for min pointe. Skiftet er allerede i gang.

Vil jeg forstå, hvordan jeg får mest ud af min investeringsportefølje, har jeg ikke ligefrem lyst til at tilmelde mig et onlinekursus i investering. Jeg starter bare en samtale med en AI-platform, der allerede kender mit vidensniveau, mine mål, de spørgsmål jeg har stillet før, og de beslutninger jeg står og skal træffe. Den kan forklare et begreb, teste om jeg har forstået det, skrue op eller ned for sværhedsgraden, lave eksempler ud fra min faktiske portefølje og skifte spor fuldstændig, når mine behov ændrer sig undervejs. Det er adaptiv læring, mere adaptiv end de fleste produkter, der kalder sig adaptive læringsplatforme, for den tilpasser ikke bare min rejse gennem et fast kursus. Den bygger læringsoplevelsen op omkring mig, mens vi går.

En typisk edtech-platform starter uden nogen viden om mig overhovedet. Jeg skal oprette endnu en konto, igennem endnu et onboarding-forløb og måske tage en diagnostisk test. Og selv da ved den kun, hvad jeg har lavet inde i netop den platform. Min foretrukne AI-platform ved måske allerede, hvad jeg har læst, skrevet, bygget og kæmpet med på tværs af hele mit liv og arbejde. Hvordan skal en almindelig læringsplatform overhovedet konkurrere med det?

Chegg er ikke historien. Chegg er advarslen

Chegg er nok det tydeligste offentlige eksempel på, hvad der sker, når de lærende holder op med at søge et dedikeret uddannelsesprodukt for svar. På sit højeste under pandemien var Chegg værdisat til næsten 15 milliarder dollar. I juli 2026 var værdien faldet til omkring 100 millioner dollar. Behovet for hjælp forsvandt ikke, studerende havde stadig brug for forklaringer og assistance. Det, der forsvandt, var grunden til at betale Chegg for adgang til dem.

Generativ AI kan i dag besvare mange af de samme spørgsmål, folk plejede at gå til Chegg med, øjeblikkeligt. Chegg har selv erkendt, at generativ AI og Googles AI Overviews har skåret i trafik og abonnementer, og virksomheden har efterfølgende skåret voldsomt i medarbejderstaben. Chegg er børsnoteret, så vi kan se kollapset i aktiekursen med det blotte øje.

Drev jeg en edtech-fond, ville jeg være langt mere bekymret for de virksomheder, hvis nedtur endnu ikke er synlig. Private virksomheder står stadig noteret til de værdiansættelser, tidligere finansieringsrunder satte. Institutionelle kontrakter kan tage år at løbe ud, og universiteter og store virksomheder flytter sig langsomt. Så omsætningen kan sagtens fortsætte, selv mens selve produktets eksistensberettigelse forsvinder under det.

En portefølje kan se sund ud på papiret længe efter, at investeringstesen er holdt op med at holde vand. At klistre en AI-assistent oven på den eksisterende platform løser ikke det problem. Spørgsmålet er ikke, om en edtech-virksomhed bruger AI, det gør stort set alle nu. Spørgsmålet er, hvorfor en lærende overhovedet skulle logge ind på platformen frem for bare at lære inde i det AI-miljø, vedkommende alligevel bruger til alt andet.

Hvad kan overleve?

Jeg tror ikke, al edtech forsvinder. Men jeg tror, at 90-95 % af dagens edtech-virksomheder på sigt bliver lukket, opkøbt eller reelt irrelevante, og det sker inden for de næste 2-3 år.

De, der overlever, skal tilbyde noget, en AI-platform ikke bare kan genskabe ved at undervise dig. Det kan være et ægte fællesskab, en stærk følelse af identitet og tilhørsforhold, adgang til mennesker man rent faktisk gerne vil lære med eller af, et bevis der har reel værdi, en fysisk eller social oplevelse, eller et stærkt idémæssigt fundament, der gør deltagelsen til andet og mere end informationsindsamling.

Folk søger ikke ind på Harvard, fordi Harvard sidder på information, man ikke kan finde andre steder. De søger ind på grund af det, Harvard repræsenterer, hvem der ellers er der, og hvad medlemskabet åbner døre til. Den samme skelnen kommer i stigende grad til at afgøre, hvilke læringsvirksomheder der overlever. Information er ikke nok længere. Indhold er ikke nok. Personalisering er ikke nok. End ikke adaptiv læring er nok, for de generelle AI-platforme kan formentlig gøre det bedre, med mere kontekst, uden at bede den lærende starte forfra.

Teknologien kom endelig. Nu risikerer den at ødelægge den kategori, den blev bygget til at redde

I over ti år troede jeg, at adaptiv læring var uddannelsens fremtid. Det havde jeg ret i. Men jeg gik ud fra, at adaptiv læring ville forvandle læringsplatformene. Nu tror jeg snarere, at den kommer til at erstatte mange af dem. Den største trussel mod edtech er ikke en anden edtech-virksomhed med flere funktioner. Truslen er, at læring bliver til bare én adfærd blandt mange inde i et langt større AI-miljø.

Samtidig opdager de organisationer, der plejer at købe læringsteknologi, at de sagtens kan bygge det meste selv. Det efterlader den traditionelle edtech-branche presset fra begge sider. Kunderne har måske ikke længere brug for at købe softwaren, og de lærende har måske ikke længere brug for at besøge platformen overhovedet. Så når aktionærer, medarbejdere og andre spørger, hvor jeg tror branchen er på vej hen, kan jeg ikke give dem det beroligende svar, de nok forventer af en, der har brugt mere end ti år på at bygge i den.

Jeg tror, uddannelse går ind i en af de mest spændende perioder i sin historie. Men bestod min investeringsportefølje af virksomheder bygget op om at levere den gennem separate softwareplatforme, ville jeg skide i bukserne lige nu.

Serie · Sådan forsvinder 90 % af edtech

En serie i fire dele om pengene, markedet og modellerne, der gjorde en hel branche til noget andet.

Del 1 af 4Hvis jeg drev en edtech-fond, ville jeg skide i bukserne

Del 2 af 4 · Næste i serienDin edtech-investor vil have et åbent forhold

Del 3 af 4Edtech sender stadig dvd'er med posten

Del 4 af 4Når din kunde bliver din konkurrent: Kunden fandt en genvej til softwarefabrikken

Læs hele serien →Beslægtede essays

AI & edtech · 18. juli 2026Hvis jeg ser én til pushe “PedTech”, kaster jeg opPædagogikken har aldrig manglet som idé. Det svære har været at bygge en branche, hvor økonomien lod pædagogikken blive stående i centrum, når investorerne bankede på.

AI & samfund · 8. september 2026Elon Musk købte den forkerte ejendom til alt-appenMusk købte Twitter for at bygge alt-appen. Men mens han samlede opmærksomhed, samlede OpenAI intentioner, og det kan være den mest værdifulde adresse på internettet.

AI & samfund · 16. august 2026AI åd internettet. Nu skal den afgøre, hvem der er menneskeVandmærkning skal gøre syntetisk indhold synligt. Den kan også gøre AI-selskaber til dem, vi beder om at attestere menneskeligt forfatterskab.

← Tilbage til blogOm Sahra-JosephineOplægBook mig til et oplæg

---

### https://sahra-josephine.com/da/blog/pedtech

← Blog

AI & edtech·18. juli 2026Hvis jeg ser én til pushe “PedTech”, kaster jeg op

Denne tekst er maskinoversat fra engelsk. Nuancer kan gå tabt undervejs, så vi anbefaler at læse originalen. Læs den engelske original

Pædagogikken har aldrig manglet som idé. Det svære har altid været noget helt andet: at bygge en branche, hvor økonomien tillod pædagogikken at blive stående i centrum, dengang investeringskronerne bankede på.

Jeg er tilsyneladende nået dertil i min karriere, hvor idéer, vi diskuterede som nybrud for over ti år siden, nu dukker op igen under nye navne og bliver solgt til mig som revolutionerende opdagelser. Når jeg følger debatten om “PedTech”, kan jeg ikke lade være med at tænke på kejserens nye klæder. Alle står og beundrer det nye tøj, mens jeg står i hjørnet og overvejer, om man overhovedet må sige højt, at vi har set det før.

Det seneste sæt tøj hedder “PedTech”. Argumentet lyder umiddelbart fornuftigt nok: Vi har fokuseret for meget på teknologien i uddannelsessektoren og bør designe forfra ud fra pædagogikken, så vi sikrer bedre læring og reelt designer for læringsudbytte. Start med læringsproblemet. Forstå den lærende. Støt læreren op. Lad være med at digitalisere indhold, bare for digitaliseringens skyld. Byg teknologien op om den måde, mennesker rent faktisk lærer på.

Ja, selvfølgelig. Det sagde vi for ti år siden, og andre sagde det længe før, jeg overhovedet trådte ind i branchen. Lærerne sagde det, før de fleste edtech-virksomheder eksisterede. Forskere har bygget hele karrierer på det. Instruktionsdesignere havde metoder til det, og edtech-stiftere skrev det direkte ind i deres pitch decks. Så når “pædagogik først” bliver præsenteret som korrektionen på en forrige generation af edtech, synes jeg faktisk, den underliggende historiefortælling er en smule fornærmende.

Jeg har aldrig ansat nogen til den pædagogiske side af CanopyLAB, som ikke brændte dybt for læring. Problemet var ikke, at en hel generation glemte pædagogikken. Problemet er langt mere ubehageligt end som så: rigtig mange mennesker, der virkelig brændte for pædagogikken, byggede virksomheder inden i et økonomisk system, som gang på gang gjorde det svært at holde pædagogikken forrest. Den forskel betyder noget.

Ingen valgte edtech, fordi det var den nemme vej til penge

Der er ved at opstå en underlig karikatur af den forrige generation af edtech. Vi skulle åbenbart have været en flok teknolognørder, der blev så begejstrede for software, at ingen huskede at spørge, om eleverne rent faktisk lærte noget. Jeg genkender ikke den branche.

Selvfølgelig var der opportunister. Der var dårlige produkter, stiftere der med tiden gik mere op i vækst end i resultater, og investorer der først og fremmest kiggede på afkastet. Sådan er det med enhver teknologikategori. Men uddannelsesteknologi har aldrig været det oplagte sted at søge hen, hvis man udelukkende vil maksimere det finansielle afkast.

Da jeg fortalte mine venner, at jeg forlod universitetsverdenen for at blive stifter i techbranchen, blev de begejstrede. Da jeg fortalte dem, at det var en edtech-virksomhed, jeg byggede, mindede de mig i stedet om, at jeg havde arbejdet med rådgivning inden akademia, og at det nok gav bedre mening, både økonomisk og karrieremæssigt, bare at vende tilbage dertil.

Havde jeg dengang vidst, hvad jeg ved nu, ville jeg have overvejet grundigt, om jeg overhovedet skulle kaste mig ud i edtech. Salgscyklusserne mod institutioner er lange. Indkøbsprocesserne er opslidende. Budgetterne strammes. Lærerne er overbebyrdede. Skoler er politiske størrelser. Universiteter flytter sig langsomt. Implementering koster dyrt. Uddannelsessystemer er fragmenterede på tværs af lande, kommuner, institutioner, læreplaner, sprog og lovgivning. Den, der bruger produktet, er sjældent den, der køber det, og køberen bestemmer måske slet ikke, om nogen ender med at bruge det. Selv når produktet fungerer upåklageligt, kan det være overraskende svært at bevise, at nogen rent faktisk lærte noget af den grund.

Der har altid været lettere steder at sælge software. Rigtig mange gik ind i uddannelsesteknologi, netop fordi de brændte for uddannelse. Lærere blev stiftere. Forskere blev stiftere. Forældre blev stiftere. Folk, der selv havde oplevet elendige uddannelsessystemer på egen krop, blev stiftere. Folk, der troede adgangen kunne blive bedre, blev stiftere. Folk, der mente teknologi kunne give den lærende noget, systemet ikke kunne, blev stiftere. Mange af os var idealistiske nok til at tro, at bedre pædagogik og bedre teknologi kunne forstærke hinanden. Det forhindrede os ikke i at bygge en branche med alvorlige problemer, men det ændrer diagnosen.

At brænde for pædagogikken reddede os ikke

Det er her, jeg mener, PedTech-debatten er ved at ramme forbi. Gode intentioner giver ikke automatisk gode systemer. En stifter kan brænde for lærerne og alligevel ende med at bygge software, ingen lærere bruger. En virksomhed kan starte med fremragende læringsforskning og ende med udelukkende at optimere gennemførelsesprocenter. En instruktionsdesigner kan skabe et flot læringsforløb og så opdage, at køberen kræver 147 funktioner i udbudsmaterialet. Et team kan tro fuldt og fast på den lærendes selvstændighed og alligevel opdage, at den største kunde vil have obligatoriske forløb, rapportering og administratorkontrol.

En stifter kan ønske at lette lærernes arbejdsbyrde og alligevel blive tvunget til endnu et dashboard, fordi den, der underskriver kontrakten, skal bruge noget målbart at vise sin chef. En virksomhed kan gå oprigtigt op i læringsudbytte og stadig opdage, at logins, klik, sete minutter og gennemførte kurser er langt lettere at proppe ind i en kvartalsrapport.

Ingen behøver være ond, for at det går sådan. Incitamenterne skal blot pege en anelse forskellige steder hen, for over tilstrækkeligt mange år bliver små retningsforskelle til vidt forskellige destinationer. Derfor synes jeg, “vi skal sætte pædagogikken først” er utilstrækkeligt som svar. Mange forsøgte netop det. Det mere brugbare spørgsmål er, hvorfor det så ofte ikke lykkedes at holde den der.

Måske var fejlen strukturel snarere end filosofisk

Forestil dig, at hele den forrige generation havde fået bedre rådgivning: pædagogik først, diagnosticér problemet, forstå den lærende, støt læreren, og mål resultaterne. Ville branchen nødvendigvis se markant anderledes ud i dag? Det tror jeg faktisk ikke.

Til syvende og sidst skal nogen betale. Det kan være en kommune, et universitet, en virksomhed, et ministerium, en forælder eller en investor, og hver eneste af dem indfører begrænsninger, der har meget lidt at gøre med læringsteori. Kommunen vil have integration. Universitetet vil have compliance. Virksomheden vil have rapportering. Indkøbsafdelingen vil have sikkerhedsdokumentation. Investoren vil have vækst. Bestyrelsen vil have tilbagevendende omsætning. Stifteren vil bare have runway nok til at overleve endnu et år.

Så træder det flot diagnosticerede pædagogiske problem ind i en kommerciel maskine, og den maskine laver om på tingene. Måske kræver den bedste løsning stor involvering fra lærerne, hvilket gør implementeringen dyr. Måske skal den lærende kun bruge produktet en sjælden gang, hvilket får fastholdelsen til at se elendig ud. Måske er den rigtige løsning fem gode læringsøjeblikke frem for halvtreds timers indhold, hvilket får kataloget til at virke tyndt. Måske er den mest pædagogisk ansvarlige konklusion, at institutionen slet ikke har brug for endnu en platform, hvilket gør salgsmødet akavet.

Konflikten stod aldrig bare mellem pædagogik og teknologi. Den stod mellem pædagogik og økonomien omkring teknologien.

Sætter man pædagogikken virkelig først, findes der somme tider ingen softwarevirksomhed

Det er den test, der interesserer mig. Forestil dig, at nogen diagnosticerer et uddannelsesproblem til perfektion. De forstår den lærende, læreren, konteksten og det, evidensen peger på virker. Efter alt det opdager de, at den bedste løsning er en lærervejledning, tre øvelser til print og en ugentlig samtale. Fantastisk. Er det PedTech? Nok ikke. Er det pædagogisk solidt? Det kan sagtens være tilfældet.

Forestil dig så et andet problem, hvor den bedste løsning er en WhatsApp-gruppe, en arbejdsbog og adgang til en mentor af kød og blod. Igen: måske fremragende uddannelse, men stadig ikke ligefrem en spændende venturefinansieret softwarevirksomhed. Her bliver “pædagogik først” for alvor mere krævende, end det lyder ved første øjekast. Pædagogikken skylder os ikke software. Den skylder os hverken tilbagevendende omsætning eller ventureafkast, og den er fuldstændig ligeglad med, om løsningen har bruttomarginer over 80 procent.

Sætter vi for alvor pædagogikken først, må vi acceptere, at teknologien nogle gange kommer i anden række, nogle gange slet ikke kommer med, og engang imellem ikke bør komme overhovedet. Det giver et vanskeligt problem for en branche, hvis forretningsmodel forudsætter, at teknologien forbliver et sted tæt på centrum.

Og timingen for rebrandingen er værd at bemærke

Jeg ville nok være mindre irriteret over PedTech, hvis det dukkede op i et vakuum, men det gør det jo ikke. Edtech gennemgår lige nu en identitetskrise som kategori, på nøjagtig samme tidspunkt hvor mange opdager, at de måske aldrig rigtig var edtech-virksomheder eller edtech-investorer til at begynde med.

Investorer, der historisk har været forbundet med uddannelse, beskriver i stigende grad sig selv ud fra bredere territorier. Jeg har tidligere argumenteret for, at edtech bliver til læring, læring bliver til kompetencer, kompetencer bliver til workforce development, og workforce development ender som human capital, karrieremobilitet, produktivitet, AI-adoption eller menneskeligt potentiale. Virksomhederne gør det samme. Uddannelsesvirksomheder bliver til kompetencevirksomheder. Læringsvirksomheder bliver til workforce-platforme. Virksomheder inden for økonomisk dannelse bliver til penge-apps. Kodeprodukter bliver til AI-adoptionsprodukter.

Mange af de forandringer giver skam rigtig god strategisk mening. Markeder flytter sig, teknologien ændrer sig, og virksomheder bør udvikle sig med tiden. Jeg har ingen særlig tilknytning til at holde en virksomhed fast i en kategori, der ikke længere beskriver, hvad den rent faktisk laver. Men timingen er værd at lægge mærke til.

Edtech har været igennem kollapset venturekapital, skuffende resultater på de offentlige markeder, lav udnyttelse af software, og nu konkurrence fra generelle AI-produkter, der kan levere mange af de ting, som før krævede en specialiseret platform. Præcis i det øjeblik opdager overraskende mange, at de ikke længere hører hjemme i edtech. Og midt i den identitetskrise dukker PedTech så op.

Måske er det en genuint ny disciplin, og det er jeg faktisk åben over for at blive overbevist om. Men når en branche i krise pludselig finder et nyt navn til et af sine ældste principper, synes jeg altså, vi har lov til at kigge nærmere på tøjet, før vi klapper ad kejseren.

Man kan ikke rebrande sig ud af sine incitamenter

Lad os sige, at samtlige edtech-virksomheder skiftede navn til PedTech i morgen. Hvad ville reelt være anderledes mandag morgen? Skolens indkøbsafdeling ville sende det samme udbud ud. Erhvervskunden ville bede om nøjagtig den samme rapportering. Læreren ville stadig have det samme antal timer i døgnet. Investoren ville stadig forvente et afkast, bestyrelsen ville stadig følge den tilbagevendende omsætning, salgsteamet ville stadig have et budget, og kunden ville stadig forvente implementering.

Intet vigtigt ville have ændret sig, og derfor interesserer det mig langt mere, om PedTech repræsenterer en anden økonomisk model og leveringsmodel, end om det repræsenterer en anden filosofi. Begynder kunder at betale for dokumenterede resultater i stedet for blot adgang, er det interessant. Får lærerne rent faktisk mere tid i stedet for endnu en administrativ opgave, er det interessant. Bliver evidens central i indkøb, bliver produkterne mindre fordi mindre er pædagogisk bedre, eller bliver investorer trygge ved at finansiere virksomheder, hvis uddannelsesmæssige værdi ikke naturligt oversættes til klassisk SaaS-økonomi, så begynder vi at tale om strukturel forandring.

De forandringer ville betyde noget, fordi de rykker ved incitamenterne omkring pædagogikken. At skrive “pædagogik først” på et slide gør det ikke.

Den forrige generation fortjener kritik. Der var for mange platforme, for mange ubrugte licenser, for meget indhold, for lidt evidens og alt for mange engagementstal, der blev forvekslet med læring. Mange produkter krævede enorme adfærdsændringer af lærere, der i forvejen var kørt trætte, og alt for meget kapital gik til virksomheder, hvis økonomi formentlig aldrig kunne bære det. Nogle af os var med til at bygge den branche, og vi bør turde indrømme, hvor vi tog fejl.

Men hvis mennesker med gode intentioner, relevant faglighed og et oprigtigt engagement i uddannelse alligevel endte med produkter, der havde lav udbredelse, svag evidens og svær økonomi, så er det interessante spørgsmål ikke, om de huskede at sætte pædagogikken først. Det er, hvorfor det så ofte ikke var nok bare at brænde for den. Måske belønner indkøbsprocesser funktioner mere pålideligt end resultater. Måske skubber venturekapital uddannelsesvirksomheder mod vækstmodeller, der slet ikke passer til uddannelse. Måske tilskynder SaaS til gentagen brug, selv når god læring ikke kræver, at man bruger softwaren igen og igen. Måske værdsætter institutionelle købere og de enkelte lærende fundamentalt forskellige ting. Måske er nogle uddannelsesproblemer simpelthen ikke gode muligheder for venturefinansieret software. De spørgsmål er betydeligt sværere at svare på end at opfinde en ny kategori, og det er netop derfor, de er mere brugbare.

En ny kategori bør forklare noget nyt

Jeg er ikke imod nye begreber som sådan. Brugbare kategorier hjælper os med at skelne mellem ting, der tidligere lignede hinanden, fordi noget reelt har ændret sig. SaaS beskrev en genuint anderledes leverings- og forretningsmodel, mens generativ AI beskriver en genuint anderledes teknologisk formåen. En brugbar kategori bør altså kunne fortælle os, hvad der er ændret.

Så sig mig: Hvad ændrede sig egentlig med PedTech? Hvad kan en virksomhed i dag, som en pædagogisk drevet uddannelsesvirksomhed ikke kunne for ti år siden? Hvilken ny leveringsmodel findes der? Hvilken ny økonomisk model findes der? Hvilket incitament er blevet rettet op på? Hvilken institutionel begrænsning er forsvundet? Hvilken evidensstandard har ændret sig, og hvad har ændret sig i køberens adfærd?

Kan PedTech svare på de spørgsmål, lytter jeg mere end gerne. Men er svaret bare, at vi skal starte med pædagogikken, diagnosticere læringsproblemet, forstå den lærende og støtte lærerne, må jeg desværre sige: du har ikke opfundet en kategori. Du har beskrevet det, som god edtech hele tiden forsøgte at blive.

Den forskel betyder noget. For hvis den forrige generation primært fejlede, fordi den havde den forkerte filosofi, er løsningen dejlig enkel: Lær alle den rigtige filosofi, giv den et nyt navn, og prøv igen. Var fejlen derimod strukturel, bliver arbejdet foran os langt hårdere. Så må vi gentænke, hvad der finansieres, hvad der købes, hvad der måles, hvordan læring leveres, hvor softwaren egentlig hører hjemme, og om enhver værdifuld uddannelsesindsats overhovedet behøver at blive en skalerbar techvirksomhed.

Det ville være en reel nulstilling. PedTech er indtil videre bare et ord. Vi bør naturligvis sætte pædagogikken først, og det burde vi have gjort både i går og i morgen, men vi skal ikke forveksle genopdagelsen af et gammelt princip med at ændre det system, der gang på gang har gjort princippet svært at følge i praksis.

Den samtale, der er værd at tage, handler ikke om EdTech over for PedTech, hvilket ord der får det næste konferencespor, eller hvem der pludselig har opdaget, at lærere betyder noget. Spørgsmålet er, om vi endelig kan bygge en økonomisk model og leveringsmodel, hvor det, alle påstår at sætte først, rent faktisk får lov at blive der.

Kan PedTech det, må I gerne klæde det på, som I vil, og jeg skal med glæde indrømme, at kejseren har fået ny garderobe. Men har vi bare genopdaget pædagogikken og givet den et konferencespor, må nogen på et tidspunkt sige det højt: Vi har set det tøj før. Bliver det præsenteret for mig som et revolutionerende nyt sæt tøj én gang til, kaster jeg måske virkelig op.

Beslægtede essays

AI & edtech · 5. september 2026Når din kunde bliver din konkurrent: Kunden fandt en genvej til softwarefabrikkenFjerde del af serien om, hvordan 90 % af edtech forsvinder. AI gør vejen fra kontoret til softwarefabrikken kortere, og kunden behøver ikke længere købe det færdige produkt.

AI & edtech · 25. juli 2026Hvis jeg drev en edtech-fond, ville jeg skide i bukserneEfter ti år i edtech tror jeg, at 90-95 % af dagens edtech-virksomheder om 2-3 år enten er lukket, opkøbt eller helt irrelevante.

AI & edtech · 13. august 2026Edtech sender stadig dvd'er med postenHistorien om, hvorfor 90 % af edtech forsvinder: folk holdt aldrig op med at lære, de skiftede bare måde, og den forskel kan vælte store dele af branchen.

← Blog

---