We start in May 2020. Thomson Reuters sued ROSS Intelligence, a startup building an AI legal research tool. The complaint alleged that ROSS had used a third-party platform called LegalEase to generate bulk research memos derived from Westlaw headnotes, then trained its AI system on the results. Three years later, and with the ROSS lawsuit still pending, Thomson Reuters agreed to buy Casetext, whose CoCounsel product used GPT-4 to help lawyers research, review, and draft faster, for $650 million in cash.

Those moves looked contradictory only if you thought Thomson Reuters was fighting legal AI itself. Instead, what it was defending was the editorial layer that made Westlaw commercially valuable in the first place: the headnotes, the Key Number classification system, and the structured workflow built on top of public law. Judicial opinions are public domain. The way West Publishing organized, annotated, and connected those opinions over more than a century is not. That is the distinction the ROSS litigation rested on, and it is the strategic thesis on which Thomson Reuters has staked its future.

In February 2025, the Delaware court granted Thomson Reuters partial summary judgment in the ROSS case. The ruling was fact-specific, and appellate review is now underway. It did not establish a universal rule about AI training and fair use. But it reinforced, in a specific set of circumstances, that curated legal editorial content has protectable commercial value even when the underlying judicial decisions do not. For Thomson Reuters, the abstraction underlying the distinction between public law and private editorial infrastructure is the business model.

This article traces how that business model was built and where AI is taking it, from a $200 newspaper purchase in 1934 to a company that reported $7.5 billion in revenue and a 39.2 percent adjusted EBITDA margin in fiscal year 2025, then watched its stock fall roughly sixty percent from a July 2025 all-time high. The story is about two things: the economics of trusted legal content, and what happens to those economics when machines can read.

Roy Thomson bought the Timmins Daily Press in northern Ontario in 1934 with $200 of his own money. He was forty, had failed at farming, auto parts, and radio, and the town was a mining settlement. Nothing about the purchase suggested much. But Thomson had an instinct for a specific kind of business: one where the cost of creating content is high, distribution is cheap, and the local market supports exactly one provider. He accumulated newspapers across Canada and then Britain, acquiring The Scotsman in 1953. Each deal followed the same pattern. Buy the franchise, cut costs, reinvest the cash flow.

The pivot from newspapers to professional information came in 1996, when Thomson Corporation acquired West Publishing for $3.43 billion, roughly four times revenue. The Department of Justice reviewed the deal for antitrust concerns and required Thomson to divest more than fifty publications before approving it. The price and the scrutiny reflected what Thomson was actually buying: the National Reporter System, the foundational compilation of American case law, and the West Key Number System, a taxonomy of more than 100,000 legal concepts that organized how lawyers found and cited decisions.

West did not own the court opinions, but built and claimed the editorial apparatus: headnotes summarizing each case's holdings, the Key Number classifications linking cases across doctrines, and the annotations connecting one decision to another. Over decades, that editorial layer became so deeply embedded in legal practice that the distinction between the law and West's organization of the law blurred. Lawyers from the start of law school are exposed to this system and it is as much a part of the assumed background of practicing law as a Bloomberg terminal would be for a finance professional.

In 2008, Thomson Corporation merged with Reuters Group for $16.6 billion. Reuters brought financial terminals, news, and a global footprint. Thomson brought legal and tax content and the professional workflow tools that generated recurring subscription revenue. The combined company had $12.6 billion in revenue and operations in more than a hundred countries. On paper, it was the world's largest professional information company. In practice, it was a conglomerate whose two halves had different economics, different competitive positions, and different growth trajectories.

Jim Smith became CEO in January 2012 and spent much of the next several years trying to stabilize Thomson Reuters' largest business, Financial & Risk. F&R generated $6.1 billion of the company's $11.3 billion in 2017 revenue, making it just over half of the group. But its growth profile was inconsistent: consolidated organic revenue was flat in 2014 and returned to positive growth in 2015. F&R struggled to grow sustainably in a market where Bloomberg was the benchmark rival, and the legal and tax businesses, which were performing well, remained invisible inside the conglomerate's blended results.

On January 30, 2018, Thomson Reuters agreed to sell 55 percent of F&R to a Blackstone-led consortium in a deal valuing the business at about $20 billion and expected to deliver approximately $17 billion in gross proceeds. The cleanest way to describe the aftermath is that Thomson Reuters deliberately shrank its scope. Pre-divestiture consolidated revenue had been $11.3 billion in 2017, while 2018 continuing-operations revenue was $5.5 billion on a restated 2017 continuing-operations base of $5.3 billion.

Thomson Reuters ultimately returned $10 billion of the proceeds to shareholders: $6.5 billion via a tender offer, $2.3 billion via a return of capital, and $1.2 billion via NCIB repurchases. The chronology ran from a $500 million buyback announced in May 2018 to an additional $1 billion in June, then in August the company said it expected to return $9 billion to $10 billion overall and formally launched the tender on August 28. Bloomberg reported the shares rose as much as seven percent intraday on the August announcement.

Post-close, the adjusted EBITDA margin dropped to 24.8 percent, the trough for the post-divestiture company. The segments that remained were already running at margins far above the blended average. Legal Professionals would reach 47.3 percent in fiscal year 2025. Tax, Audit & Accounting Professionals hit 47.1 percent.

The Refinitiv postscript validated the deal structure. Blackstone's consortium agreed definitive terms with the London Stock Exchange Group on August 1, 2019, at roughly $27 billion in enterprise value. When that transaction closed on January 29, 2021, Thomson Reuters indirectly owned approximately 82.5 million LSEG shares, worth roughly $9.8 billion based on LSEG's closing price the prior day. The board had sold a structurally disadvantaged business at a fair price and kept a meaningful share of the upside.

Steve Hasker was announced as Smith's successor on February 25, 2020, effective March 15. His background was McKinsey (1998 to 2009), then COO of Nielsen, then a stint at TPG Capital. In February 2021, he launched a two-year Change Program with expected investment of $500 million to $600 million. By the end of 2022, the program had delivered $540 million of run-rate operating expense savings. Over roughly the same period, Westlaw Edge ACV penetration rose from 46 percent as of September 2020 to about 70 percent by late 2022. The margin recovery was real but not perfectly linear: adjusted EBITDA margin moved from 24.8 percent in 2018 to 25.3 percent in 2019, jumped to 33.0 percent in 2020, dipped to 31.0 percent in 2021 as Change Program costs flowed through, then rose to 35.1 percent in 2022.

The revenue rebuild, in exact figures looked as follows: $5.501 billion in 2018, $5.906 billion in 2019, $5.984 billion in 2020, $6.348 billion in 2021, $6.627 billion in 2022, $6.794 billion in 2023, $7.258 billion in 2024, and $7.476 billion in 2025. The post-divestiture company generated $2.936 billion of adjusted EBITDA at a 39.2 percent margin in fiscal year 2025, versus $3.437 billion at 30.3 percent for the 2017 consolidated company. While revenue is a third lower, the margin on every dollar is meaningfully higher.

In March 2023, OpenAI reported that GPT-4 had passed a simulated version of the Uniform Bar Examination, though later commentary questioned the most aggressive percentile framing. Steve Hasker later described the announcement as an oh, wow moment. On June 22, 2023, Judge P. Kevin Castel sanctioned two lawyers in Mata v. Avianca after their filing cited nonexistent cases generated by ChatGPT, imposing a $5,000 joint penalty.

Together, those events exposed the core tension in legal AI that we have seen play out in recent years: benchmark performance could look impressive, but unchecked generative output could still be professionally dangerous. This is a tension we have since seen repeatedly. In a profession where a fabricated citation can lead to sanctions or malpractice claims, the difference matters.

Four days after the Mata sanctions order, Thomson Reuters announced its $650 million acquisition of Casetext. At that point, CoCounsel was Casetext's GPT-4-powered legal assistant. It was not yet a Thomson Reuters product grounded in Westlaw's editorial layer. Thomson Reuters' strategy was to pair generative AI with its own trusted content and workflows in stages: in November 2023, it launched Westlaw Precision AI-Assisted Research drawing on Thomson Reuters content; in early 2024, it described the integration of CoCounsel with Westlaw and Practical Law as the turning point. The distinction matters because the ROSS case is about training inputs, while the CoCounsel product story is about retrieval and answer-grounding. Those are different legal and technical questions.

Hasker framed the broader approach as a Build, Buy, Partner strategy. In a 2025 Semafor interview, he said Thomson Reuters had spent about $500 million on AI-driven product development between 2023 and mid-2025. The company's earlier disclosures were more incremental: more than $100 million per year in generative AI investment as of November 2023, growing to more than $200 million in 2024. On the acquisition side, the major deals were Casetext ($650 million), Pagero (approximately $800 million for global e-invoicing), and SafeSend (approximately $600 million for tax automation), along with smaller transactions. (For another day is whether the thesis that grounded AI eliminates hallucination risk is even correct. There have been recent cases where lawyers cited CoCounsel as the source of hallucinated results.)

By year-end 2025, Thomson Reuters reported that 28 percent of annualized contract value came from generative-AI-enabled products, up from 18 percent at year-end 2024. The Big Three segments delivered 9 percent organic growth for full-year 2025. These are large numbers for a company that was growing at one percent six years ago, but they are a bit ambiguous. Thomson Reuters counts ACV from any product that is GenAI-enabled, whether or not the customer is actively using the generative features. In the Q4 2024 earnings call, management noted that one point of the sequential ACV increase came from divestitures being removed from the denominator, with the rest driven mostly by Westlaw Precision and CoCounsel adoption. Until the company discloses how much of the ACV represents genuinely new revenue versus reclassified existing subscriptions, the metric is directionally useful but not conclusive.

The ROSS litigation sits alongside these product developments, and it is important to state precisely what the February 2025 ruling did and did not do. On February 11, 2025, Judge Stephanos Bibas granted Thomson Reuters partial summary judgment, holding that ROSS's use of Westlaw-derived headnotes, obtained through LegalEase Bulk Memos, was not fair use on the specific facts before him. The court found infringement of 2,243 headnotes. The opinion expressly noted that ROSS's tool was not generative AI and that only non-generative AI was before the court. Judge Bibas subsequently granted ROSS interlocutory appeal and a stay pending appeal, so appellate review is underway.

For Thomson Reuters' strategic position, the ruling matters less as settled law and more as a signal of direction. If the appellate courts affirm, competitors building AI legal research tools will face a practical choice: license comparable editorial rights if available, or build a substitute editorial layer from scratch. The court explicitly noted that ROSS could have created the material itself or hired LegalEase to create it without infringing. If the ruling is reversed or narrowed, the editorial moat still exists as a product-quality advantage, but the legal barrier around it weakens.

Using split-adjusted U.S. closing prices, Thomson Reuters rose from the mid-$40s after the Refinitiv transaction closed in October 2018 to a closing high of $210.94 on July 14, 2025, before falling back to the high-$80s by late March 2026 — a decline of roughly fifty-nine percent from peak to trough. The decline coincided with a broader repricing of AI-adjacent equities. Whether Thomson Reuters underperformed its professional-information peer set by more than that sector-wide move requires a comparison I have not done here, but the magnitude of the drawdown suggests company-specific concerns beyond sector rotation.

The bull case rests on fundamentals that have not changed. Revenue is eighty-one percent recurring. Management reports a ninety-two percent retention rate, though the metric's scope — whether it applies uniformly across all segments or reflects a blended figure weighted toward Legal Professionals — is not fully specified in public filings. The Big Three segments grew nine percent organically in fiscal year 2025. Free cash flow was $1.95 billion. The company has raised its dividend for thirty-three consecutive years and increased it ten percent in fiscal 2025 to $2.62 per share annualized. At a late-March 2026 share price around $87 and roughly 445 million shares outstanding, the free cash flow yield is approximately five percent on a company growing organically at seven percent. Management's 2026 guidance calls for 7.5 to 8 percent organic revenue growth, adjusted EBITDA margin expansion of about 100 basis points to approximately 40.2 percent, and free cash flow of roughly $2.1 billion.

The company estimates its Big Three segments address a $31 billion served market, a figure introduced in the 2025 annual report. Separately, the 2024 Investor Day presented a $26 billion vended market alongside an $84 billion broader total addressable market that included generative AI expansion opportunities. These are different numbers serving different rhetorical purposes at different points in time. The $31 billion figure represents the current professional information and workflow market where Thomson Reuters competes directly. The $84 billion figure includes markets the company would need AI to enter. Treating them interchangeably, as some sell-side coverage has, overstates the near-term revenue opportunity.

As of its most recent disclosure, the company describes approximately $11 billion in remaining capital deployment capacity through 2028, which includes balance sheet capacity and expected free cash flow generation. Previous investor materials referenced $10 billion through 2027. How that capital is deployed will shape the next five years of returns. Casetext was acquired for $650 million, but Thomson Reuters has not disclosed Casetext's standalone revenue at the time of the deal, so the implied valuation multiple is not publicly verifiable. If subsequent acquisitions are struck at high multiples during a period of AI hype, the returns on invested capital could lag the cost of capital even if the products succeed.

The bear case centers on three expense and commitment trends that are visible in the filings.

The first is the software-related expense trajectory. Thomson Reuters reported $721 million in software amortization for fiscal year 2025, up from $618 million the prior year. But the 2026 guidance figure of $890 to $910 million is for software depreciation and amortization combined, a different and broader measure, versus $832 million on that same basis in 2025. Those are still meaningful year-over-year increases that reflect the rapid capitalization of AI-era software development and, in some cases, shorter assumed useful lives. The distinction matters for which profit line is affected: adjusted EBITDA excludes software amortization entirely, so these increases do not compress the adjusted EBITDA margin that management highlights. They do, however, weigh on operating profit and adjusted earnings per share.

The second is the purchase obligation surge. Future unconditional purchase obligations rose to $1.96 billion in fiscal year 2025 from $883 million the prior year. The filing describes these as commitments for technology, outsourcing, and other services. They may include AI-related cloud infrastructure, but the filing does not break out the components, so it is not possible to say with certainty how much of the increase is AI-driven. On a $1.95 billion free cash flow base, a billion dollars in incremental forward commitments narrows the margin for error regardless of their specific composition.

The third is the Reuters News anomaly. Management said Reuters had $33 million of one-time generative AI content licensing revenue in fiscal year 2024 and $13 million in 2025, while the segment's full-year EBITDA margin fell from 23.6 to 20.4 percent. The decline suggests that AI content licensing deals with model providers were front-loaded or lumpy rather than recurring, complicating the narrative that Thomson Reuters' content will generate a steady stream of licensing income from AI companies. Reuters contributes less than twelve percent of total revenue, so the direct financial impact is modest. The signaling effect is not.

Goodwill on the balance sheet stands at $7.9 billion against $7.5 billion in revenue. The ratio reflects the cumulative price paid for West Publishing, Casetext, and dozens of smaller acquisitions. It is not unusual for a serial acquirer of professional information assets. But it means that a sustained thesis break, particularly one that impairs the value of Westlaw's editorial infrastructure, would create balance sheet consequences well beyond the income statement.

The simplest version of the bear case is this: an open-source legal AI model, trained on publicly available court decisions rather than proprietary editorial content, that performs adequately for routine research. Thomson Reuters' competitive position depends on the premise that professional-grade AI requires professional-grade editorial content. If a free or low-cost alternative proves good enough for a meaningful share of the market, the moat narrows. The ROSS ruling, even if upheld on appeal, protects Thomson Reuters from unauthorized copying of its editorial work. It does not protect the company from a competitor that builds a different editorial layer from scratch, or from one that persuades enough lawyers that the editorial layer is unnecessary.

Thomson Reuters has shaped how lawyers organize legal knowledge for more than a century. The West Key Number System gave the profession a shared classification framework for case law, and generations of lawyers have internalized it so thoroughly that it functions less like a product and more like professional infrastructure. That embeddedness is the moat. The question is how long it holds.

CoCounsel is Thomson Reuters' bet that the moat extends into the AI era. Instead of navigating key numbers and headnotes, a lawyer describes a question in natural language and receives synthesized analysis drawn from the same underlying editorial content. But natural language search on Westlaw is not new. Westlaw Edge pushed AI-assisted search as early as 2018. I was still litigating then, and I found the structured approach — terms and connectors, key numbers, jurisdiction filters — more reliable. I left practice in 2018 and have not used CoCounsel, so I cannot speak to whether the current product is materially better. The point is adoption friction, and that friction predates the current AI cycle.

That friction matters because the AI investment thesis depends on adoption. Thomson Reuters reports that twenty-eight percent of annualized contract value now comes from generative-AI-enabled products, but as discussed in the prior section, the company has not disclosed how much of that represents new revenue versus existing subscriptions reclassified after a generative feature was added. Reported hallucination incidents continue to surface in legal filings, and each one reinforces the caution that practicing lawyers already feel. If CoCounsel's AI-assisted research proves trustworthy enough to change behavior at scale, the investment pays off. If lawyers treat it the way many treated natural language search in 2018 — as an interesting option they mostly ignore — then the roughly $2.6 billion Hasker has cited in AI-era acquisitions and the approximately $500 million in product development spending he described in a 2025 Semafor interview are chasing adoption that may not come fast enough.

The deeper risk is structural. The West Key Number System's value depends on lawyers organizing their work around it. If AI changes how legal research is conducted at a fundamental level — if lawyers stop navigating taxonomies and start querying models that synthesize across multiple sources without reference to headnotes or key numbers — then the editorial layer that took a century to build may stop functioning as a moat. Thomson Reuters' differentiation is not just raw text. It is editorial structure, citator logic, trusted content, and workflow integration. But if enough of the market concludes that an adequate alternative can be built on public opinions alone, the moat shrinks even if it does not disappear. The ROSS ruling protects Thomson Reuters from unauthorized copying. It does not protect the company from a future in which the classification system matters less because lawyers interact with legal knowledge differently.

The 2025 annual report introduces agentic AI as the company's preferred framing, supplementing the generative AI language of prior years. Thomson Reuters describes agentic systems as planning and executing multistep legal workflows with minimal human input, while still requiring human oversight, interpretation, and validation. For a profession built on the premise that a licensed human exercises judgment at every step, this raises practical questions even though existing ethics rules already reach much of the territory. ABA Formal Opinion 512 and New York City Bar Opinion 2024-5 both apply duties of competence, confidentiality, supervision, candor, communication, and fees to lawyers' use of generative AI. The harder question is narrower: whether those existing duties adequately resolve the practical challenges of semi-autonomous legal workflows that take actions, not just produce drafts. Thomson Reuters is building these systems while the profession is still working through the answer.

Roy Thomson's $200 newspaper in 1934 was a bet that in a market with one provider of organized local information, the provider sets the terms. Ninety years later, his company is making a version of the same bet: that curated, structured legal editorial content will remain essential even as AI reshapes how lawyers work. The bet may prove right. But the headwinds are real — on adoption, on pricing dynamics, and on whether the editorial moat itself survives the transition to a fundamentally different way of interacting with legal knowledge. The stock market, as of March 2026, appears to be pricing in the possibility that it does not.