Sahra-Josephine Hjorth
Back to blog

← Blog

AI & Society

Your Dashboard Will Be the Last to Know: Thomson Reuters Fell Nearly 18% in a Day Over an AI Tool While Its Revenue Grew

When an AI company releases a tool that does what you sell, investors can take billions off your value in a day while your sales are still growing. Your sales will only show whether they were right months later, when customers cancel, so a board that wants to know first has to ask its customers how much of the work they already do without it.

Portrait of Sahra-Josephine Hjorth with long hair, lit by pink dappled light against a patterned wall

The Five Cs of Disappearance

Every month for half a century, someone in an office in New York opened fifty-one newspapers, one from each of fifty-one American cities, turned to the back pages and counted the small ads. They skipped the big display ads for cars and department stores and counted only the narrow columns of short lines where employers said they were looking for a typist, a welder or a nurse. The office added them up and published a single number.

To anyone outside economics, counting small ads every month would have looked like a strange hobby. Inside economics, it was one of the most watched figures in the country. It was the Help-Wanted Advertising Index, published by the Conference Board, a business research group. When it climbed, employers were hiring. When it fell, a slowdown was usually on its way, and economists read it the way a doctor reads a pulse.

In August 2005, the index fell to a record low. Anyone reading it alone would have concluded that American employers had stopped hiring. Yet unemployment that month was 4.9 percent, low by any standard. The jobs were there. The ads had moved online, and the Conference Board knew it, because a few months earlier it had started a second count, of job ads posted on the internet. That August, the two counts told opposite stories. The newspaper index hit bottom, while the online count found 2.13 million new job ads, more than in June. For about three years the Conference Board published both side by side, so anyone could watch the old one fall while the new one showed where the hiring had gone. In 2008 it retired the newspaper index.

I kept thinking about that index while I worked on the study behind this series. In the industries from my study that disappeared, the old investors sold only after the customers had started to go, so of the Five Cs of disappearance, Capital, the money, was usually the last to move. With AI, the order has flipped. Investors now sell within days of an AI company releasing a tool, long before any cancellation shows up in the company's sales. That leaves boards in an odd position. The people who own the company are the first to bet on whether its customers are leaving, and the company itself finds out last, from a dashboard that only moves when a contract ends.

I see this from the other side. My studio builds software for companies that they would previously have bought from a vendor. Some of the projects my own studio builds are exactly what this essay is about. In a software vendor's records, each of them is just a renewal that quietly did not happen.

Investors knocked nearly 18 percent off Thomson Reuters in a day while its sales were growing

Thomson Reuters is best known for its news agency, but most of its money comes from lawyers, accountants and tax advisers. They pay for Westlaw, its legal research service, and for software that helps them draft, review and check documents. On Friday, January 30, 2026, Anthropic released a set of add-ons for its AI assistant, and one of them was for lawyers. It could review a contract, sort through confidentiality agreements and check documents against compliance rules, which is a large part of what lawyers pay Thomson Reuters to help them do. When investors caught up with it the following Tuesday, February 3, Thomson Reuters' shares fell nearly 18 percent during trading.

Two days later, Thomson Reuters reported its results for 2025. Its sales had grown 7 percent, not counting acquisitions, and sales in its Legal Professionals segment, the part that sells to law firms, had grown 8 percent. Nothing in its own numbers showed lawyers leaving.

The same thing happened again in May. On the morning of May 5, Thomson Reuters reported a solid first quarter, and its shares went up. Later that day, Anthropic announced AI tools for financial work, and Thomson Reuters' shares turned around and fell as much as 5.1 percent before closing slightly lower.

So far, the sales have not followed the share price. In the third quarter of 2026, Thomson Reuters' sales grew 10 percent, and its Legal Professionals segment grew 9 percent. Yet on the day it reported those numbers, November 3, its shares fell about 7 percent, because investors worried about what its own spending on AI would do to its profits.

Figure 1

Thomson Reuters' shares fell on Anthropic's announcements while its own sales kept growing

How to read it. The four days, in order. Pink bars hang down from the line and show how far Thomson Reuters' shares fell on Anthropic's news. Orange bars rise from the line and show the sales growth Thomson Reuters itself reported, not counting acquisitions. Both pink figures are the steepest fall during the day. On May 5 the shares closed only slightly lower. The two kinds of number measure different things and share a scale only to show their direction.

Sources: Reuters; Thomson Reuters, February 2026; Bloomberg; Thomson Reuters, November 2026.

This is also where the help-wanted index stops being a perfect comparison. The index was a false alarm. It fell while the jobs were fine, because the jobs had only changed address. Rising sales do not prove that Thomson Reuters' investors were wrong in the same way. A share price is what investors expect the company to earn in the future, and they may be expecting lower prices, thinner margins or the cost of building its own AI long before any lawyer cancels. Rising sales can also hide the move itself. Thomson Reuters bought Casetext, maker of the AI legal assistant CoCounsel, for $650 million in 2023, so some of the lawyers it still counts may already be doing their work with AI, only AI that Thomson Reuters sells. Sales make a different mistake. They stay calm while customers are already leaving, because most customers leave at renewal, months after they have made up their minds. A board looking at a falling share price and rising sales is looking at two numbers that can each mislead it, and neither can tell it whether its customers are actually going.

An AI company's announcement is written to sell, and only your own customers can check it

In What the 6 Ds of Exponential Technology Don't Show, I wrote that the detailed record of where customers go stays with the company they go to. What the AI company publishes instead is the headline, because the AI company is selling to the same customers and raising money from the same investors. A launch announcement is a sales pitch, and that is exactly why investors react to it, and why the claims in it tend to run ahead of what customers are actually doing.

The strongest objection to everything I have said so far is that investors simply overreact, and there is a well-known case that seems to prove it. On February 28, 2024, Klarna, a Swedish payments company, announced that its AI assistant had handled 2.3 million customer conversations in its first month, the work of about 700 full-time agents. Klarna is the kind of company that hires call centers. The same day, shares in Teleperformance, the largest call-center company in the world, fell as much as 29 percent. A year later, Klarna's chief executive, Sebastian Siemiatkowski, said the company had focused too much on cost, and that "what you end up having is lower quality".

That objection is half right. Klarna does not prove that AI will leave call centers alone. It proves that an announcement tells you neither how far a customer will go nor how well the new way will work. A share price that falls on one is an expectation, formed by people who have read a sales pitch and cannot ask your customers what they will do. Investors do try to check. Analysts call a dozen law firms or a handful of a call-center company's clients. But a dozen calls is a small sample, and the company has something no investor has: every customer, every contract and every conversation about renewal.

Almost no company uses those for this question. A board looks at its sales, which arrive late, and its share price, which is somebody else's unchecked expectation, and it has nothing of its own in between. The one count a company can still own is what its own customers tell it, and the Conference Board's lesson is to start that count before the old number falls.

A board can learn what its customers are doing before they cancel

The first thing to count is the announcements themselves, because they are public and nobody inside most companies reads them as news about their own customers. When an AI company releases a tool that does part of what you sell, someone should write down what it does and which of your customers it is aimed at. When one of your large customers tells its investors that it is automating the work it used to buy from you, as Klarna did, that is news about your sales, published before your sales show it. Customers also post job ads for the engineers who will build their own tools, which makes those ads a help-wanted index of their own. Someone simply has to collect all of this every month, the way someone in New York once went through the newspapers.

The second thing is the check that only the company can make, and it has to happen before customers leave, because the ones who have already left rarely explain why. Most companies ask their customers at renewal whether they are satisfied. A satisfied customer can still be doing half the job somewhere else. The better question is what share of the work the customer now does without you, and with what. The answers come in three stages, and it helps to keep them apart. First, customers try something else alongside you. Then they move real pieces of work to it. Only after that do they buy less from you. A trial is not yet a loss, but the second stage is, even though it will not show in your sales until the third. Ask at every renewal, and ask the prospects who decided not to buy what they use instead.

The third thing is where the answers go. Put them on the same page as sales at every board meeting, give one person the responsibility for keeping them, and keep both numbers long enough to see whether they move apart. If your sales keep rising while the share of the work your customers do without you rises too, your two counts are telling opposite stories, as the Conference Board's did in August 2005. And when the share price falls on someone else's announcement, the board will have something of its own to put next to it.

Figure 2

Put what your customers are doing next to your sales, because the last line is one investors cannot get

Your sales
What most boards read today
Saleswhat customers paid you this quarter
Sales growththe same, without acquisitions
Renewalshow many customers stayed
Your customers
What they are doing instead
AI tools released that do part of what you sellfrom the AI companies' own announcements
Customers saying they will do the work themselvestheir press releases and reports to investors
Customers hiring engineers to build their own toolstheir job ads, a help-wanted index of their own
How much of the work each customer now does without you, and with whatasked at every renewal, and of prospects who did not buy

How to read it. One board page, with two columns side by side. The left column is what most boards already read, and it moves only after customers have cancelled. The right column collects what customers are doing instead, mostly from public sources that investors read too. The pink line is the one investors cannot get: what your own customers tell you at renewal. Keep both columns until you can see whether they move apart.

Your dashboard is still counting newspapers

When the help-wanted index fell to its record low, the Conference Board could have defended its number. Every ad it counted was real, and every newspaper it read was printed. Instead, it trusted the count it had already started where the ads had gone, and let the old one go. A company's dashboard is in the same position today. Every sale on it is real, and it counts only what customers still buy from you, not the work they have started doing elsewhere. In 2005 the old number raised a false alarm while the jobs were fine. Today the old number stays calm while the alarm comes from investors who have read someone else's announcement and sold before anyone has checked. The company is the only one that can check, by asking the people who pay it how much of the work they now do elsewhere. A company that does not ask is where a reader of the newspaper index was in August 2005, holding a real number and the wrong picture, except that this time the number says nothing has moved.

Under every essay in this series is the Iceman Test: eleven questions about your own industry that show which of the Five Cs are already moving. It is one way to start counting where your customers are going.

The Five Cs of Disappearance · Self-test

Is your business the iceman?

Industries rarely die because a better rival wins. They disappear because customers quietly move, and in a study of 30 industries, an industry's own numbers were almost always the last place it showed. Eleven questions, about four minutes. Answer for one thing you are paid for, not the whole company.

For example: fix cars, translate documents, sell accounting software, cater weddings.

Sources

Link to this essay
Follow me