Harvey Built Legal AI for the Work the Next Model Still Cannot Do

On Training Data, Sequoia's Pat Grady introduces Harvey as the canonical example of an application-layer company built on foundation models. The US legal market, he says, is $400 billion—the same size as the global cloud markets. A few years earlier, startups in this category were mocked as wrappers on GPT. Winston Weinberg and co-founder Gabe took the opposite view: the value would be created in the messiness of real-world problems, giving lawyers superpowers rather than replacing them.

Harvey was founded at the end of July 2022, months before ChatGPT. Asked what gave him conviction that there would be value between foundation models and customers, Weinberg starts with how messy these industries are. The largest misconception about so-called GPT-wrapper companies, he says, was that many of them had almost no delta between what the foundation model already provided and what a small company could add. Harvey's first advantage was an almost stubborn emphasis on citations: in legal work it matters that you can cite line by line, and that those citations are accurate.

If citations were the whole company, he admits, Harvey would just be a wrapper and it would go away. The ambition is different: partner with a trillion-dollar industry that is messy, full of datasets, and full of specialized workflows, and actually transform it. Models, in this view, will not automate entire industries in one shot. They will serve as the groundwork, unlocking pieces of a process as they improve.

When OpenAI released Deep Research, Weinberg's first reaction was not that capital-markets lawyers had been replaced. It was that studying markets and public data could slot into a few steps of a hundred-step process and make that process better. Whenever a foundation model gets better at something, or ships a new feature, it unlocks part of the TAM Harvey can build on top of. That is the opposite of a thin wrapper: the application keeps absorbing model progress into workflows the model cannot run alone.

The product method Weinberg says Harvey has gotten most right—and found hardest to manage—is expand, then collapse. If models worked perfectly and humans communicated perfectly, the best interface would be email, or even a neural link: no UI at all. That is not how it works, and even if the models could chain every step of a large merger like Microsoft and Activision, a user cannot send "merge please" and walk away. The human side of the workflow cannot absorb a one-shot.

So Harvey expands by building specialized, start-to-finish systems for high-value use cases that a chat box cannot cover. Case-law research, for example, needs retrieval over cases, comparison among them, and synthesis of this matter's facts against the law—three systems, not one prompt. Then those workflows collapse back into one service. Upload a share purchase agreement and Harvey can ask whether you want to extract the reps and warranties, summarize the document, or run another workflow the user never has to know was built separately. The commercial tension is real: seats have to apply to as many lawyers as possible, while specialized agentic workflows are often great for a securities attorney and irrelevant to everyone else. Collapse is how those vertical systems get packaged as a shared service.

Every workflow, Weinberg says, is three questions: what does the user want, what context do we need, and is this right? Intent is routing and follow-up questions. Context is mostly retrieval—separate systems for internal and external documents, plus routing that decides when to search which. "Is this right?" comes back to citations, and to data the models do not have. Models do not know what "market" is. Market for one private-equity firm's LBO terms is different from market across private equity, which is different from general M&A, and Harvey has to retrieve those datasets and compare them. Having those firms as customers helps.

The bigger problem is that process data for these tasks does not exist on the internet. How you do disclosure schedules, or what "market" means, is not sitting on Reddit. Harvey hires domain experts who sit down and say these are the steps they would take, then chains models on top and does task-specific fine-tuning where the model still gaps. The old hope that feeding every legal document into a model would teach it to practice law is, he says, like reading the entire case load and every law-school textbook and then being dropped into a profession. Evaluation is the other expensive pit: you hire mid-level people to judge the work, because if juniors could evaluate it they would already be more senior, and he argues that 20 to 30 percent of firm revenue is spent doing exactly that.

Weinberg does not think the end state is typing "merge company A and B" into a chat box anytime soon—not because he is bearish on models, but because that interface does not let the user exercise enough judgment on a high-stakes workflow. The agent tasks Harvey talks about cost hundreds of thousands of dollars. Legal is a good market for LLMs for two blunt reasons: the industry is text-based, and tokens are extremely valuable. In a 50-page merger agreement, every piece of a word cost a fortune to produce.

Asked whether Harvey sells software or work, Weinberg says so far, software. Early clients were law firms and professional-service firms such as PwC; enterprises have followed. The product is splitting in two: a productivity suite with lawyers in the loop at all times, whose ROI is hours saved each week, and specialized workflows that complete part of the work start to finish, which is closer to selling the work itself. Harvey is building those with firms under revenue-split agreements, combining the firm's domain expertise with Harvey's tech, then letting the firm sell it to its own clients. The company is transitioning from seats to selling work, and he expects to see a lot of that this year.

Inside the building, that shift is a bet. Horizontal features raise value for the whole user base and dilute the chance of failure. A specialized workflow is zero-one: huge if it sells, worthless if it does not. Harvey's hedge is to let law firms do discovery—joint projects with a large company and its counsel to test whether the work is repeatable, whether the tech can do it, and whether the in-house team will accept AI on that use case. On the product side, it is building about 15 horizontal "AI patterns"—case-law research, regulatory research, clause extraction—and chaining them like a Ford factory line.

Externally, the collision is with a business model that has not changed in a century. Harvey is working with a firm almost a hundred years old, famous for advising on King Edward VIII's abdication, that is now willing to bet on changing how it makes money. The billable hour makes efficiency software a hard sell: there are only so many hours in a day unless you invent the 25th. The opening move is to take work firms currently do at a loss—side-letter compliance to win private-equity business, for example—turn it into software, charge a flat fee, and help the firm take market share without eating the cost. If model reasoning froze today, Weinberg still thinks Harvey could push penetration well above the small percentage points AI currently has in legal, because the bottleneck is evaluation, process, and data, not reasoning.

The origin story goes back to early 2022. Weinberg and Gabe had GPT-3 access and were doing chain-of-thought prompting on legal questions before the industry had a name for it—breaking problems into process steps instead of one static prompt. They cold-emailed the results to OpenAI general counsel Jason Kwon, who wrote back that he had no idea the models were this good at legal. Other people, Weinberg thinks, were making one model call and calling it a day. That process data is now an unfair advantage because Harvey's customers are some of the best firms in the world, a contrarian choice: lots of firms wanted in, and Harvey said no in order to focus on the prestigious ones.

The product reason was that models would keep getting better, so you want to build systems the next generation still cannot do—complex international-merger work, the most defensible layer. The go-to-market reason is that in professional services, prestige amplifies trust. Earn the trust of a few large firms and the rest of them, plus everyone downstream, plus their clients, will trust you. Going straight to enterprise looked tempting. Enterprises had no reason to believe a startup could actually build these systems.

Trust came from things that do not scale. Weinberg says there is no excuse for an AI company not to do hyper-personalized demos—it used to be important and is now paramount. He used a partner's recent matter, then asked argumentative lawyers to fight with the output until they were reading every word Harvey produced. The product side of behavior change is killing the blank page: the landing screen is buttons, now customized by years of practice and practice area, so time-to-value is as short as possible. Customer success hired lawyers who can say they have been in the same shoes.

Sonya Huang puts the standard objection: hallucinations are a feature in creative industries and a bug in legal. Weinberg pushes back. A lot of what lawyers do is creative; litigation would not exist if ten simple rules mapped every fact pattern, and fitting facts to the boundaries of the rules is highly creative work—he was a litigator. Accuracy still ranks first. Firm hierarchy helps: a junior drafts, a second-year reviews, a fifth-year reviews, a partner reviews. Minimum viable quality can be lower for law firms than for in-house teams, because someone is always in the loop. For in-house, sell specialized tasks at extremely high accuracy, where fine-tuning and evaluation are easier because the surface area is smaller.

Looking five to ten years out, Weinberg thinks the job of the lawyer returns to what it was fifty years ago: an adviser who looks around corners. The chief legal officer title is relatively new, and most CLOs he meets are business drivers, not the person who only says no. Lower-end work gets commoditized and automated; high-level strategic work gets more valuable. That is good for young lawyers. Nobody goes to law school to sit in a data room labeling documents for ten years and maybe try a case once. They want to give advice, help people, or win, and it is closer to professional athletes: you get better from hands-on strategic experience, not from the data room.

He is now a law-firm client himself, and the annoyance is specific: top billable rates for low-end work such as checking whether a change-of-control clause was triggered. For high-end strategy—how to buy this business, how to restructure this part of the org—he would pay more than the best lawyers on earth currently charge. The delta between a junior associate and the best partner is only three or four times, which does not make sense to him. His prediction is fixed fees for the commoditized parts of a deal or a litigation, and more expensive pricing for insight and strategic advice.

Access to justice is the part of the mission he cares about most. The average US lawyer costs $352 an hour, so almost no one can afford one, and even if every lawyer gave twenty hours a week to that work the gap would not close. Two old rules sit in the way: unauthorized practice of law, and the ban on non-lawyer equity in firms, which has cut traditional financing out of the legal sphere. Utah, Arizona, and some other states are dropping those rules or running sandboxes, and Weinberg says it has been going well. The overlooked piece is that most people do not know when their rights have been violated—housing, junk fees, someone in authority doing something illegal. AI's first job may be education at scale, not only delivering the service.

Zooming out to the AI market, cheaper inference is always good. The o-series models mattered more than that: they unlocked Harvey's product strategy for the next six months to a year. Many use cases had been stuck at working 70 percent of the time, and Weinberg had not expected to build those systems in the next year on GPT-series models alone. What o-series is good at is orchestrating a plan and executing it. Harvey can already extract from EDGAR, case law, and internal documents; the missing piece was combining those fragments into the correct work product, and that is what the o-series unlocked.

If they switched jobs, Weinberg jokes, Sam Altman would raise more money at Harvey. At OpenAI, Weinberg would double down on capturing the consumer zeitgeist and productionizing the experience, not just model scores—almost none of his non-Valley friends know any AI tool except ChatGPT. The partnership has stayed strong because Harvey keeps pushing back: OpenAI says a new model wins on benchmarks, Harvey says it does not on theirs. Application-layer companies have to bet on providers, but Harvey chains different models for different pieces, and Microsoft still has to live with customers who think Claude is better at some jobs.

The threat to Harvey, if the opportunity is a trillion-dollar industry plus access to justice, is not moving fast enough. Weinberg tells the team they are living through the most compressed timeline in human history. Speed compounds: you have to test everything, watch every customer and every subprocessor, and the scariest mistake is moving too slowly and missing something massive. If steps 12 and 13 of a feature are done and the model cannot do step 13 until someone ships an unlock, that unlock has to go into the product immediately. Saying you will try it eventually is the failure mode.

In the lightning round, he has not taken a day off in nearly three years, and he is mixed about that badge of honor. Obsession matches compressed timelines; leadership still has to change as the company went from about 40 people at the start of last year to 260. Most hiring mistakes came from not understanding what the role actually did, so he still believes founders should do a bit of every job—though he did it too long. The prior he has updated most is "teach, not do": at the first sign of friction he jumps in, which is faster and keeps other people from learning. On luck versus skill, luck is timing and whether you have the option and leverage to apply skill. He has quadrupled down on young talent because their skill is the ability to adjust to luck and seize it.