Yesterday Google Cloud launched Gemini Enterprise for Legal, with Cleary Gottlieb, Freshfields, Weil and Williams & Connolly as its first firms. Google shipped a financial services version the same afternoon and says healthcare is next, so treat this as a template rather than a legal story.
Five days before that, Harvey released Tenet, its first model. Not built from scratch... they took Kimi K3, an open-weight base, post-trained it with Fireworks on around 150 Nvidia B300 chips over two months, and the post-trained weights sit with Harvey. It beats frontier models on Harvey's own contracts benchmark.
Put the two together and you get the odd shape of this market right now. The model companies are building legal products, and the biggest legal product company is building models.
Fourth in, not first
Google is the fourth big technology company to move into legal this year.
Anthropic shipped legal plugins for Cowork at the end of January, then formally launched Claude for Legal on 12 May... twelve practice-area plugins, more than eighty agents, over twenty connectors into the systems firms already run. Bloomberg Law's read on the launch was that Anthropic is positioning itself as a direct provider of legal technology in its own right, well beyond its original role as a model maker.
On 1 June, Jason Boehmig, co-founder of Ironclad, joined OpenAI to lead its legal vertical. His own description of the job was building AGI for law. Seven weeks later Willkie announced a firmwide rollout of ChatGPT Enterprise. Microsoft has had a legal agent inside Word since the spring. And now Google, with a finance twin in the same box.
The reason is boring and it fits in one survey number. 69% of legal professionals now personally use general-purpose AI platforms for work, up from 31% a year earlier. Lawyers didn't wait for legal tools. They opened ChatGPT and Claude and got on with it, and the labs are less invading the legal market than following their own users into it. The same migration is underway in finance, accounting and consulting.
Harvey's answer
Harvey's write-up for Tenet lists two goals: frontier legal intelligence built on open weights, and systems that let law firms build and own their own specialised models.
The second one is the business. When Kirkland committed 500 million dollars to a platform it refuses to license, the argument was that off-the-shelf tools raise the floor for everyone... they can't make you better than the firm across the street. Harvey has productised that argument for every firm without 500 million to spend.
And they needed a move, because their position had turned awkward. Claude for Legal integrates Harvey. Gemini Enterprise for Legal integrates Harvey. The most valuable company in legal AI is now listed as a connector inside two of its own suppliers' products. Two months on 150 chips is what answering that costs.
On Harvey's diligence benchmark, where one task can mean working through 80 million tokens of documents, no baseline model passed more than 43.8% of the criteria; one of Harvey's post-trained versions reached 60.1%. Document review across 10,000 files at a time came in at roughly a tenth of the cost per cell, with better citations. Search over a firm's own knowledge got 90% cheaper per query.
Every services market has a floor below which a matter isn't worth taking. Drop the cost tenfold and the floor moves, and it will move in accounting, insurance and claims for exactly the same reason it moved here.
What's left to defend
Step back and nobody is in their old lane anymore. Anthropic co-founded Ode, an actual standalone services firm, with Blackstone, Hellman & Friedman and Goldman Sachs, because enterprise demand for Claude was outpacing any single delivery model (its CFO's words).
McKinsey runs 25,000 agents alongside 40,000 humans and says about a third of revenue now comes from underwriting outcomes rather than billing advice. Goodwin built Regina, its own proprietary operating system. Accenture bought Faculty in January and made its founder CTO. Even inside Google's legal launch, the agents come from Eudia, which runs its own law firm through Arizona's alternative business structure, and Deloitte, with Accenture and KPMG among the named implementation partners.
Labs doing services, consultancies doing product, law firms doing software, a law firm supplying a hyperscaler. A year ago any one of those would have been its own story.
So what can't be copied? Harvey just proved the model can be... take an open-weight base, post-train for two months, beat the frontier in your domain at a fraction of the cost. If Harvey can, so can Thomson Reuters, so can Intuit, so can a large firm. Workflows can be copied too, since Google packaged legal work as an industry template and shipped financial services the same day. Firm knowledge is genuinely proprietary, but Harvey and Google are both now selling ways to make it usable, and whoever owns the client relationship decides who gets access to it.
What remains is the client and the liability. Who the client calls, and who carries the blame when the work is wrong.
My call
This is very good for firms and very hard for pure tools. A firm can now buy frontier capability from four directions and play them off against each other. A tool company has to explain why it isn't a feature of something bigger.
Three predictions. Microsoft and AWS ship their own industry packages within a year, because once two hyperscalers and two labs have packaged a profession the rest can't leave it alone. At least one large firm announces it trained its own model rather than licensing one... Kirkland is already spending 500 million dollars building its own platform, and Harvey has just shown a version of it can be done in two months of compute. And the tools layer keeps consolidating. Harvey and Legora have bought eight companies between them this year, and being a connector inside somebody else's platform is a weaker hand than it looked in January.
Eight editions in, the pattern hasn't changed. The value ends up with whoever owns the work and carries the risk of getting it wrong, not with whoever builds the technology. This month, four of the biggest names in technology arrived in one profession, the leading specialist answered by building its own model, and not one of them got any closer to owning the client.
Subscribe and every edition lands in your inbox. If you believe my read is wrong, reply. I read everything, and the sharpest replies shape the next edition.
And if you're thinking about selling a professional services firm, or buying one... that's exactly what I do. I run Eilla, an AI-native M&A advisory. Message me.
See you next time.