Harvey's strategic evolution

We've both had people say disparaging things to us about Harvey — that it's an overvalued ChatGPT wrapper, that $11 billion for a legal AI chatbot is Silicon Valley hype at its most absurd. We understand the criticism, but our own reading of what Harvey is actually doing — what the thesis behind it is — is quite different.

Anthea builds and orchestrates AI agents for complex decision-making and reads Harvey as a technology company making platform, infrastructure and ecosystem moves. David reads Harvey as a move within the century-long evolution of how legal services are organized and delivered. We come at this from different directions, and the combination matters.

Let's start with a caveat. Neither of us has tried Harvey, so we can't vouch for it as a product. Harvey is an enterprise legal product — unlike ChatGPT or Claude, you can't sign up and test it for $20 a month. Everybody's heard of Harvey, but almost no one outside the firms that have adopted it has actually used it. People don't know whether it's a souped-up legal research tool, a way of applying the baseline models to law in a more protected environment, or something else entirely. The opacity isn't secrecy — it's a structural feature of the enterprise sales model. And partly, people are dismissive of Harvey because they can't form a mental model of what it does.

There's a second group of dismissers too. Some of them know the baseline models well and believe that ChatGPT, Claude, and their successors are improving so fast that you won't need a specialized, domain-specific application built on top of them. Every time a new model comes out — each one more capable than the last — this view gets reinforced. Why pay for Harvey when the general model can already do the work? It's a serious argument, and one we'll come back to: foundation model companies have begun moving into specific verticals themselves, which puts pressure on any layer built on top of them.

Others see the more immediate threat as Harvey's direct competitors. Legora — a Stockholm-based AI-native legal platform — is the most prominent: growing fast, picking up major US and European firms as customers, and pursuing a similar enterprise-platform strategy on a different timeline. The legal AI category may not end up with a single dominant player. The more plausible picture is two or three serious platforms sharing the top of the market, each making slightly different bets about where defensibility comes from.

Both arguments have force, and the second of them — the question of whether vertical legal AI can hold its position as the underlying foundation models become more capable and more direct in their pursuit of the legal market — is one we genuinely don't know the answer to. But we also think dismissing Harvey as a mere ChatGPT wrapper is wrong, and here's why.

The Five Rings

Harvey isn't really a product in the way many people think of an AI product. We view it as having gone through a five-part transformation that positions it as something more interesting — and potentially much more consequential — than an AI tool.

Harvey's evolution is better captured as concentric circles than as sequential pivots. Each phase builds on and incorporates the previous one. Nothing gets replaced. Everything gets absorbed into an expanding structure. We've put together a visual showing the five rings with a timeline of key milestones — let's walk through them.

The visualisation at the top of this piece is interactive — click any ring to explore the detail behind that phase.

It starts with the founding story. Winston Weinberg, a litigation associate at O'Melveny & Myers, and Gabe Pereyra, who'd worked on large language models at Google DeepMind, Google Brain, and Meta, were roommates. They got early access to GPT-4 six months before ChatGPT launched and deliberately resisted the standard startup advice to pick a narrow use case — go draft NDAs, make that work really well. Their thesis was the opposite: "the models are the product." A bet on breadth. A conversational AI associate that could do many things rather than one thing perfectly. That's a key insight: because they had GPT-4 before anybody else, they recognized it had broad general potential for legal workflows, not just narrow task automation.

They also had the advantage of being incubated by a firm that was genuinely forward-thinking. Allen & Overy (now A&O Shearman) didn't just become their first customer — the firm allowed Harvey to experiment inside its walls under a mandate from the top, not just to find use cases but to work through the key challenges that were stopping people from using the technology. It was a true incubation, not just a vendor relationship. A&O rolled Harvey out to 3,500 attorneys through a top-down, enterprise-wide deal. This sort of acceptance by a major law firm was an enormous boost to the credibility of this fledgling startup. It also meant that they became aware very early on of the sorts of difficulties that would come with rolling out AI enterprise solutions in law. That's the inner ring — the broad assistant.

Then, rather than chasing new features, Harvey did something that looks boring but turned out to be highly strategic. Given its early experiences in law firms, it decided to invest in security and governance. Security infrastructure, isolated computing environments, admin controls, links to firm systems, separation of client matters. What they realized was that governance wasn't just a compliance burden. In legal, it's a competitive lever. It gets a product into firms. Once they had that structure, they could credibly say: what you do inside Harvey isn't going to get outside Harvey. It positions the company as the AI name that firms can trust. As a16z's Marc Andrusko put it: "Nobody gets fired for buying Harvey." That's the second ring — enterprise trust based on governance and security.

The third ring is where Harvey stopped being a tool and started being a platform. Legal work organises around matters — a merger, a patent prosecution, a securities litigation — and Harvey's architecture followed that reality with matter-centric workspaces. Their Workflow Builder let firms encode proprietary knowledge into structured, repeatable processes. The numbers are striking: 18,000+ custom workflows, 25,000+ custom agents. Firms weren't just using Harvey. They were building on it. Users feel much more ownership over AI tools they build and customize, though we don't know how portable these workflows would be off Harvey, so we don't know how sticky they are. This was an ecosystem and platform play that went beyond a narrow AI tool story.

The fourth ring is where Harvey stopped being a platform and started becoming infrastructure. In December 2025, Harvey launched Shared Spaces — a feature that lets law firms and their clients work in the same AI-powered workspace. Guest accounts mean firms can invite clients who aren't Harvey customers. The design partnerships are revealing: PwC and IFS working alongside law firms on M&A analysis. Gleiss Lutz and Deutsche Telekom collaborating on regulatory matters. Three types of organization — corporate buyer, consulting advisor, law firm — in the same workspace on the same matter. Harvey is positioning to become the infrastructure layer that coordinates complex deals and all the players involved in them.

Then there is the outermost ring, which is still forming. In early 2026, Harvey started acquiring. It bought Hexus, a startup that builds AI-powered product demos and guides — not a legal AI company, but an engineering talent play. This was the first of at least two acquisitions in the year, a platform company's pattern of absorbing capabilities it can't build fast enough. That's a different posture from building products. It's the behavior of an entity assembling an ecosystem.

Then Harvey embedded itself in the foundation model layer. It became a launch partner on Anthropic's Claude Marketplace — alongside Snowflake, GitLab, and Replit — and integrated with Anthropic's Model Context Protocol, so that Claude users can invoke Harvey's legal workflows from inside Claude and get results back in the same thread. Harvey also moved to a multi-model strategy — OpenAI, Anthropic, and Google — becoming model-agnostic rather than tied to any single foundation model.

This was a revealing strategic choice. LexisNexis took the opposite approach, absorbing Claude capabilities inside its own Protégé platform to reinforce its data dominance. Harvey chose to be complementary middleware — going to where the users are rather than pulling users to itself. The bet is that enterprises will standardize on Claude or another foundation model, and Harvey will be there when they arrive. The counter-bet — held by some of the dismissers from the section above — is that the foundation model companies themselves will move into the customer relationship and the legal-specific workflows, leaving Harvey squeezed from above as well as from below.

And Harvey became an investor. It partnered with The LegalTech Fund — a $110 million venture firm backed by law firms McDermott and Orrick — to co-invest in legal tech startups. Up to $2 million per startup, focused on "point solution providers" — complementary tools, not competitors. The amounts are modest. But Weinberg's framing was telling: he described the goal as developing a "marketplace model similar to approaches used by Anthropic."

That phrase — "marketplace model similar to Anthropic" — is the tell. Harvey isn't just participating in an ecosystem. It's modelling itself on how foundation model companies govern theirs. The fifth ring isn't about building better tools or even better infrastructure. It's about shaping who else builds, what gets built, and how capital flows through the legal technology landscape.

Beyond the Product Story

Now, why does this matter beyond Harvey's own business story?

Harvey hasn't gone for the lower-value automation work — the contract review tools, the document classifiers, the e-discovery platforms. Its founders figure that this work will be done by other AI native startups now and likely the models themselves as they improve. Instead, Harvey is trying to become the infrastructure through which lawyers, their clients, and other professional services firms interact in a safe, secure, and AI-enabled way — and, increasingly, the entity that shapes the ecosystem of tools and companies that plug into that infrastructure.

If you think about AI and law in isolation, you think about how AI changes the legal field — who wins and who loses within it. Lawyers have strong threat perceptions when it comes to AI. But when you zoom out, something more interesting comes into view. As we noted in our previous piece, Harvey at $11 billion is roughly ten times the largest accounting AI company. No standalone consulting AI company exists at meaningful scale. Harvey isn't just dominant in legal AI. It's dominant across professional services AI.

What if law firms are not a beachhead for AI in the legal industry? What if they are a beachhead for professional services more generally? If that were the case, would Harvey's strategy give law a chance to reposition itself as a central node in the broader AI transformation of professional services fields? If so, does just Harvey benefit, or do law and legal firms structurally benefit too?

Harvey's strategy is worth watching closely because it maps onto the actual role that elite lawyers play — not just the technical work they do, but the integrative judgment at the complex intersection of law and business and strategy and politics and public relations that we described as law's gateway position.

a16z's Andrusko identified this possibility three months before Shared Spaces launched, drawing an analogy to accounting. QuickBooks and similar products benefited enormously from accountants acting as distribution nodes — a small business and its external CPA sharing the same file, using the same tools. No legal tool had achieved the equivalent. Could legal tech get law firms and their corporate clients working together in one system?

Harvey's answer, three months later: Shared Spaces.

The reframe is this. The question isn't whether Harvey is a good AI tool. It's whether Harvey is building the infrastructure through which AI-powered professional services are delivered — with law firms as a central node in the distribution channel. A different story entirely from "legal chatbot valued at $11 billion."

The Adoption Gap

None of this is to say that Harvey and their backers are right.

One word of caution: adoption in terms of licenses is reportedly high, but actual use apparently remains relatively low. a16z's own analysis acknowledged that Harvey is "there and available" at major firms but "not heavily used yet." The simple explanation is that lawyers are technology averse and have been doing things the same way for a long time. But the deeper reason may be an identity threat.

What most lawyers think — particularly junior to mid-level lawyers — is that the basis of what they do is the ability to read and regurgitate legal documents and legal information in a technical way. That creates a double bind with AI: if the tool isn't reliable enough, it makes mistakes, and lawyers don't trust it. If the tool is too reliable, it threatens to put them out of business, and lawyers are afraid of it. Either way, they resist.

Harvey hasn't yet created the kind of pull where the tool reshapes how you work rather than slotting into how you already work — the kind of pull that is an unmistakeable sign of product-market fit at the use level rather than the purchasing level. We don't hear people raving about Harvey and using it for things well beyond law in the same way as we hear people raving about Claude Code and using it for things well outside coding.

But the picture may be more bimodal than "low usage" suggests. Harvey's own data indicates that for users who have adopted four or more product lines, daily active usage runs at 75-85 percent — numbers that rival the most engaging consumer products. The pattern isn't uniform disengagement. It's a sharp split: most lawyers barely touch the tool, but those who cross some threshold of integration become deeply embedded in it. This fits what Anthea has observed more generally about AI adoption — it tends to be bimodal rather than normally distributed, with a small group of power users whose behaviour looks radically different from the majority. We plan to explore this in a future piece, because the traits of these AI super-users and power adopters may tell us something important about the nature of the future cyborg lawyer — the professional who doesn't just use AI tools but works with AI as a genuine cognitive partner or extended mind. Understanding what distinguishes that group from the rest of the profession may matter more than any aggregate adoption statistic.

And there are structural tensions that remain genuinely unresolved. Law firms bill by the hour, and Harvey makes them faster — which means, in the short term, AI efficiency doesn't obviously help the business model. The problem runs deeper than pricing. Law firms are built like pyramids: high-end judgment and client relationships at the apex, commodity technical work done by junior people at the base. The base subsidises the apex — that's how firms make money. AI is automating the base. The technical part of what lawyers do is already being automated, and it will only accelerate as the tools get faster, more used, and more accepted. If AI hollows out the base, the economics of the whole structure are threatened.

Harvey, for its part, isn't simply waiting for the billing model to shift. It has begun experimenting with a different revenue structure entirely — moving from seat-based software pricing to revenue-share arrangements where Harvey co-builds AI-powered workflows with law firms, and those firms commercialise the output for their clients. The effect is to flip the firm's relationship with the tool: Harvey becomes something the firm sells through, not just something it pays for. And because the work being sold is professional services rather than software licences, the addressable budget is orders of magnitude larger. If it works, it's a partial answer to the billing paradox: not faster hours, but different work sold differently. But it's early, and whether this model scales beyond a handful of sophisticated firms remains to be seen.

Building for the Third Path

In our previous piece, we described three paths that technology creates for any profession: doing the same thing cheaper and faster, expanding demand through lower costs, or doing fundamentally different things. Most lawyers are still on the first path. Harvey's inner rings serve that path well — faster research, more efficient drafting, better governance.

But Harvey's outer rings — the platform, Shared Spaces, and the ecosystem play — are infrastructure for the third path: the one where law's role in professional services gets redefined, not just accelerated. Putting law firms, consulting firms, and corporate clients in the same AI-powered workspace isn't a feature for doing legal work faster. It's a feature for doing different work entirely — integrated, multi-disciplinary, with law as the coordinating node. And the fifth ring goes further still: if Harvey governs the ecosystem of tools that serve law, it doesn't just coordinate the work. It shapes the landscape in which the work happens.

The Real Question

So is Harvey an overvalued ChatGPT wrapper? We don't think that's the right question. The question is whether law's structural position — touching every major business decision, embedded in every transaction, present at every closing table — makes it the natural distribution channel for AI across professional services. If so, the "wrapper" dismissal doesn't just underestimate Harvey. It misunderstands what Harvey is trying to become.

But Harvey faces a genuine multi-track challenge. It has to provide enough value within the existing framework of what lawyers do — the cheaper-and-faster path, the governance and trust that gets it adopted — while simultaneously trying to move the best lawyers and the most forward-thinking firms into the next phase of what professional services is and where law fits into that. And it has to do both while building an ecosystem governance role that only works if the inner rings are thriving first. You can't govern an ecosystem people aren't actively using — and the adoption gap remains real.

There's also a question that cuts deeper. Harvey's evolution from product to platform to infrastructure to ecosystem governor looks, from one angle, like brilliant strategic escalation. From another, it looks like the classic pattern of a technology company moving upmarket — retreating from the hard, unresolved problem of product-market fit at the usage level into the structurally easier work of capital allocation and marketplace orchestration. Investing in startups and co-producing industry foresight is not the same as getting 100,000 lawyers to change how they work every day. Whether the outer rings can expand while the inner rings remain underutilized is the question the next few years will answer.

Update — May 2026

We drafted this essay at the end of March. Four developments in the weeks since are worth flagging because they actively pressure-test the strategic read above.

Harvey moved further into the ecosystem-governor role. On 5 May 2026, Harvey launched a library of 500-plus purpose-built legal agents, an Agent Builder that lets firms create custom agents in plain English, and a public benchmark called Legal Agent Bench (LAB) — 1,200-plus tasks across 24 practice areas, evaluated by 75,000-plus expert-written rubric criteria. The set of partners on LAB is the giveaway: Nvidia, OpenAI, Anthropic, Mistral, and DeepMind, alongside LangChain, Fireworks AI, Stanford Liftlab, and Snorkel. Owning the benchmark is a classic ecosystem-governor move — whoever defines what "good" looks like in the category shapes what everyone else builds toward. This is also the most explicit statement yet of Weinberg's framing: "The legal industry is now well past AI as an assistant and officially in the era of legal agents." If the Ring 5 ecosystem-governor reading is right, this is exactly the kind of move you'd expect — and it strengthens the platform-and-infrastructure thesis above. Harvey also disclosed an updated user base of more than 100,000 lawyers across 1,500 organisations.

Slaughter and May went firmwide with Harvey on 29 April 2026. A second Magic Circle firm at scale, supported by Harvey's Transformation Office, in addition to A&O. The credibility flywheel we describe in Ring 3 is visibly continuing.

Legora reached a $5.6 billion valuation on 30 April 2026, with Nvidia making its first legal-tech investment via NVentures. Legora has crossed $100 million in ARR and now lists Cleary Gottlieb, Linklaters, White & Case, Bird & Bird, Barclays, and HSBC among its customers — with meaningful overlap into firms that also use Harvey. The two-or-three-platforms picture we sketched in the dismissers section is taking concrete shape; the category is forming into something more like a duopoly at the top than a single dominant platform.

Freshfields signed a multi-year direct partnership with Anthropic on 23 April 2026, deploying Claude across 5,700 employees and entering a co-development program to build legal-focused AI applications. This is the first Magic Circle firm to choose a foundation-model partner over a vertical legal-AI vendor at scale. Combined with Anthropic's launch of Claude for Word in beta on 11 April 2026 — which leads with legal contract review on its product page — this is the counter-bet from the multi-model paragraph showing up in the wild. And Anthropic's late-April moves were quickly followed on 4 May by a $1.5 billion services joint venture with Blackstone, Goldman Sachs, and Hellman & Friedman — Anthropic isn't just letting firms go direct, it's setting up its own services-delivery arm — with OpenAI announcing a parallel $10 billion services venture with TPG, Brookfield, Advent, and Bain Capital the same day. For firms with the internal capability to do their own AI engineering, the intermediation layer is increasingly optional. Whether Harvey's infrastructure-and-ecosystem play continues to compound or whether the bigger firms route around it is now one of the live questions in the category.

None of these developments invalidate the strategic read above. The agents-and-LAB launch in particular sharpens it: Ring 5 is escalating, not stalling. But the duopoly and the going-direct datapoints do mean the picture has more tension in it than it did a month ago.

But Harvey's story — whatever its outcome — illuminates something important about the broader landscape. The gap between purchasing and usage isn't unique to Harvey. It's a pattern across the profession. Adoption in terms of licences is reportedly high; actual use remains relatively low. Part of the reason is structural — billing models, risk aversion, regulatory caution. But part of it is more personal: most lawyers haven't yet crossed the gap between talking about AI and working with it. That gap — what it feels like, what it takes to close it, and why leaders in particular can't afford to delegate it — is where we turn next.