Harvey vs Thomson Reuters: Is This the First Real Incumbent Kill?

Harvey reached an $11 billion valuation in March 2026 on roughly $190 million of annual recurring revenue, a multiple near 58 times. It sells directly against Thomson Reuters, which has spent well over a billion dollars defending a legal segment generating $6.8 billion a year. Legal was supposed to be the hardest vertical to disrupt, which makes this the clearest available test of whether AI-native companies actually displace incumbents.

The numbers on both sides

This comparison is usually made with one side's figures. Here are both.

Harvey reached an $11 billion valuation in March 2026 on roughly $190 million of ARR, having been valued at $8 billion three months earlier. One analysis put the resulting multiple near 58 times revenue, against the 15 to 20 times forward multiples typical of high-growth vertical software.

Independent estimates put Harvey's ARR higher by mid-year, around $300 million in May 2026, up from $195 million at the end of 2025. The company reports usage across more than 100,000 lawyers. That spans over 1,300 organisations in more than 60 countries, including roughly half the Am Law 100.

Thomson Reuters is defending a legal segment generating approximately $6.8 billion annually. It acquired Casetext for $650 million in 2023 and bundled the resulting CoCounsel product into Westlaw. By February 2026 it reported CoCounsel passing one million professional users across 107 countries.

What the incumbent actually did

The standard disruption narrative has the incumbent ignoring the threat until it is too late. That is not what happened here.

Thomson Reuters bought a leading AI-native legal assistant and integrated it into its flagship research product. It made the result available on predictable per-seat pricing, to a customer base already paying for Westlaw. It then shipped agentic capability grounded in its own content.

That is close to the optimal incumbent response. Acquire the capability, bundle it into the distribution you already own, and price it as an add-on rather than a new purchase decision.

The result is that a firm already on Westlaw faces a considerably lower barrier to adopting CoCounsel than to signing a new enterprise agreement with a challenger. Distribution beats product more often than founders like to admit, a pattern examined in the analysis of category absorption.

Why acquiring beat building here

The incumbent could have built an AI assistant internally. It bought one instead, at considerable cost, and integrated it into a product customers were already paying for.

That decision bought two things a build could not: time, and a team that had already learned what lawyers actually needed. In a category where the challenger had a three-year head start on product understanding, buying the understanding was faster than acquiring it.

The alliance that changed the competitive map

The challenger's structural weakness was content. Legal research requires a verified, citable database, and building one is a multi-decade undertaking that two companies have already completed.

Harvey addressed this through a partnership with LexisNexis, structured so that Harvey remains the interface while LexisNexis supplies the content. That arrangement removes the objection litigators and research-heavy practices raised most often.

It also creates a dependency. The challenger now relies on a company that competes with its main rival, which is a strong position while the alliance holds and a serious exposure if it does not.

For every other vendor in the category without a comparable data partnership, the alliance widened a structural content gap that is very difficult to close quickly.

Is this displacement, or is the market expanding?

This is the question the headline framing usually skips.

Legal AI venture funding reached approximately $4.3 billion across 180-plus deals in 2025, against roughly $2.1 billion for the five years from 2019 to 2023 combined.

Both the challenger and the incumbent are growing. Harvey's usage rose sharply while CoCounsel passed a million users. Neither company's growth appears to be coming primarily from the other.

The most defensible reading is that AI expanded the legal software market rather than reallocating it. Work previously done by junior lawyers, or not done at all, is now being done by software. Both vendors are selling into that expansion rather than taking share from each other.

That distinction matters for anyone applying this case to their own category. An expanding market rewards both incumbent and challenger for several years, and it produces headlines that look like disruption while both parties grow.

Displacement may still come. It has not happened yet, and describing this as an incumbent kill is currently a prediction wearing the clothes of an observation.

What the benchmarks say, and what they do not

Published benchmark comparisons put the leading vertical products within a few points of each other, with the challenger typically at the top of that narrow band.

More interesting is the comparison with general-purpose models. At least one published evaluation found a frontier general model outscoring every legal-specific platform tested, at API prices. The stated catch was verification burden, with a substantial share of frontier-model answers citing or applying law inaccurately in the same evaluation.

That is the real competitive question for the category. If a general model is more capable but requires you to build citation verification and jurisdiction governance yourself, the vertical product is selling the verification layer rather than the intelligence.

Two cautions on all benchmark figures here. Several published comparisons are produced by companies competing in the category, which is a conflict readers should weight heavily. And the challenger open-sourced its own legal agent benchmark in May 2026, with contributions from several model developers, which is a genuine contribution and also a benchmark designed by a participant.

The bear case on the challenger

Three arguments, and they are not weak.

Market size, competition and the multiple

The addressable market is narrower than the valuation implies. An enterprise-only strategy with substantial seat minimums and annual commitments structurally excludes the mid-market and small firms, which is where most of the world's two million-plus lawyers actually work.

Competition compresses margin from three directions. The incumbent bundles, a European rival competes directly on enterprise workflows with transparent pricing, and general assistants are used informally for ad hoc tasks at no incremental cost.

The multiple assumes winner-take-most. At roughly 58 times revenue, the valuation requires legal AI to consolidate around one platform. If the market fragments by firm size and jurisdiction, which is how legal services have always been organised, the multiple does not survive.

What this tells you about incumbents generally

The legal case is instructive well beyond legal, because it shows what a competent incumbent response looks like.

The fourth row is the one that does not transfer. Incumbents defending a category with no proprietary data asset have a considerably weaker hand, and the first three moves alone will not hold a position indefinitely.

Where this reading is weak

Several problems worth stating.

Harvey is private and does not publish financials. Every revenue figure here is an estimate or press report, and different sources give materially different numbers for the same period, ranging from $190 million to $300 million ARR within a few months.

The incumbent's segment revenue is not all contested. Comparing a $190 million challenger to a $6.8 billion segment overstates the overlap, since most of that segment is research subscriptions the challenger does not sell against directly.

And five years is short. Legal buying cycles are slow, partnerships shift, and the alliance that currently resolves the content problem is a commercial agreement rather than an asset. This picture could look entirely different in eighteen months.

Frequently asked questions

Where to start this week

If you buy vertical AI in any category, one question separates the durable products from the interfaces.

Ask the vendor what they own that a general model plus a competent engineer could not assemble in six months. Verified data, a compliance surface and accumulated domain logic are real answers. Prompt quality and user interface are not.

Then ask what happens to their product if their content partnership ends. A vendor that has thought about that has a strategy. A vendor that has not is renting its moat.

The same two questions work in reverse if you are building in a vertical. If your honest answer to the first is prompt quality and interface, you have identified the work for the next twelve months.

Related reading