Winston Weinberg

The through-line

Winston Weinberg's consistent argument is that legal work is far harder than technologists assume, and that this difficulty is the opportunity rather than the obstacle. "Most folks in tech do not appreciate how complex legal work is," he says. "People think what lawyers do is easy. It's not. It's incredibly complex, and there's a reason that lawyers get paid the way they do" [6]. From that premise everything else follows: models will not sweep an industry away on their own, so the value sits in the messy layer between the model and the customer, built with the industry rather than at it [3][4]. He has held this line since Harvey was dismissed as a wrapper on GPT, and the evidence he points to is that "these models weren't going to just automate entire industries away, right? They were going to change and like fundamentally serve as kind of like the groundwork for changing an entire industry" [3].

The second, less visible through-line is that his own job is a permanent rebuild. "It feels like every six months you have to like reearn your position in the market and that trickles down through the whole company where you have to reearn your position at the company, including myself" [5]. He describes the same recurring failure mode across four years, and says the fix has to be different each time, "because if it's like basically the same problem and then the fix is the same every time, that means you have not improved or learned in any way, shape, or form" [2].

On why legal work resists automation

The core reason is that the raw material is missing from the internet. "The process data for a lot of these tasks doesn't exist on the internet," he says, using disclosure schedules and the question of "what is market" as examples of knowledge that is not sitting on Reddit somewhere [3]. Harvey's answer is to "actually hire domain experts who sit down and say these are the steps that I would take," then chain models on top of that, and where a gap remains, fine-tune. His view on training is emphatic: "the best way to do fine-tuning and post- training is task specific." The idea that you could pour all the legal documents into a model and get a lawyer out he compares to reading every case and textbook and then being dropped into the profession, expecting to know how it works [3]. Evaluation is the other half of the problem and is equally expensive, because judging the work requires mid-level people rather than juniors, and evaluating junior work is itself a large share of what professional services firms do, perhaps 20 to 30% of revenue [3].

He also insists the economics of the domain are unusually favourable. The test he applies is whether an industry is text-based and how valuable a token is, and "a token is incredibly valuable in legal and professional services," since each word of a fifty-page merger agreement is expensive to produce [3]. When the models improve or a new capability ships, he reads it as territory opening up rather than a threat: deep research, for instance, immediately suggested capital markets use cases that could be dropped into one step of a hundred-step process [3]. That same complexity makes accuracy non-negotiable, which is why citations, line by line and accurate, were a resource priority from day one, and why auditability and security matter so much that Harvey's security team is unusually large relative to engineering [3][4]. He has since discussed his view that model performance is plateauing and what that means for enterprise adoption and the future of professional services [8], and the capabilities and limits of AI agents in a broader market conversation [12].

On expanding and collapsing the product

His central product doctrine is a rhythm: "at all times you have to basically expand the product and then collapse it back" [3]. In a perfect world where models and humans both communicated flawlessly, the ideal interface would be email and nothing else, but that world does not exist, and even if the models could chain every step of a merger like Activision and Microsoft, no user is going to type "merge please" and walk away [3]. So Harvey expands by building specific vertical agentic systems for high-value tasks, case law retrieval, comparing and contrasting cases, synthesising facts against precedent, then collapses them into one surface, so that uploading a share purchase agreement prompts Harvey to offer the relevant workflows without the user knowing they were built separately [3]. The tension he names is commercial: seat-based software has to serve as many users as possible, so you are always balancing whether to build for the securities attorney or for every attorney [3].

Underneath this he uses a three-part decomposition of any workflow: what does the user want and how do you extract that intent, what context is needed, and is this right. Routing and follow-up questions serve the first, retrieval across internal and external documents serves the second, and the third is where citations and the ability to retrieve and compare different market data sets earn their keep [3]. The end state he describes is not a simpler UI so much as a smarter one: the text window looks similar while the suggestion, routing and orchestration models improve, so the system behaves like a colleague who knows what you want to do and remembers what you did last time [3]. He frames it as building specialised associates alongside the partner or managing partner operating model that pulls them together [3]. The routing extends outward too, into a Word plugin that hands you off to Word when you start drafting, on the principle that "you meet the users where they are" [4]. Generalisation comes from identifying roughly fifteen AI patterns in legal, case law research, regulatory research, clause extraction, and investing heavily in those horizontal capabilities [3]. Specificity, meanwhile, is what makes measurement possible: a workflow that takes target and acquirer financials and tells you whether merger clearance is an issue in 75 countries is a closed universe with a set of regulations and facts, and can be benchmarked properly, whereas "find every single case that has ever said XYZ" cannot [4].

On partnering with the industry rather than building in a lab

He rejects the approach of declaring a vertical traditional and outdated and building everything in Silicon Valley before releasing it to run over everything [4]. Investors told him going after the hardest customers first was a horrible idea; he did it because "trust is the most important thing in professional services," and prestige is the proxy for it, so "if you earn the trust of a few of those firms, the rest of them will trust you and the rest of the firms downstream will definitely trust you, right? And their clients will trust you" [3][4]. Going straight to enterprise would have left no reason for anyone to believe Harvey could build these systems [3]. Partnering with firms like Allen & Overy gave two things at once: design partners who revealed which use cases mattered, and an education in what the industry actually cares about, namely accuracy, auditability and security [4][6].

The internal version of that education is cultural. He asks how you teach engineers what lawyers do, and then says the more important question is "how do you help them respect what lawyers do?" [4]. At around thirty people, Harvey brought in attorneys who had worked on taking Dell private to walk engineers through how they decided what a tracking stock was, and he watched the room shift to thinking the work was hard and impressive [4]. Lawyers sit with engineers, help design the systems and evaluate the outputs, and he says the answer to which parts of the product involve lawyers is all of them, even as the process has become more operationalised at scale [4].

On demos, deployment and how adoption spreads

The unscalable thing that worked was making every demo personal. Because filings are public even when work product is not, he would find a firm's most recent brief or a merger agreement from one of their deals and have Harvey attack it, telling litigators to find the holes in their own arguments [4][6]. It was risky, since early on the model would sometimes be wrong, but it solved the attention problem: "you tell a litigator that like there might be holes in this argument and they go you know but they want to prove you wrong, right? And so they really pay attention" [4][6]. That tactic got productionised into prebuilt workflows that let senior partners click one button and get a useful output without prompting, which he calls "a it's a deployment strategy" [4].

Adoption then compounds in two directions. Internal virality is real, but he thinks the more powerful force is external: firms show clients what they are doing with Harvey, the private equity fund or bank asks to collaborate in the same system, and workflows built by a law firm can be populated into that client's own instance [4]. Enterprises actively want their firms to teach them how to use AI [4]. The discovery problem is the bottleneck, since DAU over MAU for users who have touched one product feature is about half that of users who have touched three or four, and the three-to-four group is close to Slack levels [4]. Model-driven orchestration and routing are how he intends to close that gap [4].

He is also clear that adoption ran ahead in the UK first, because US banks sent letters to law firms telling them not to adopt generative AI while the UK largely did not, which is why Harvey's first two customers were in London when the company had three people [6]. Onboarding 4,000 users at that headcount meant rotations on support tickets and three days without sleep, and the waitlist that drew criticism for secrecy existed because there was no bandwidth left to demo for anyone else [6]. The way to make that many people happy at once was to go practice area by practice area and land at least one highly relevant use case for each, a closing checklist for M&A being the example, which took real domain expertise [6].

On selling work, not only seats

He describes Harvey as two products. One is a productivity suite with lawyers in the loop at all times, sold as software and justified by hours saved. The other is workflows that complete part of the work from start to finish, which is closer to selling the work itself, and Harvey is building those jointly with firms under revenue-split agreements that combine the firm's domain expertise with Harvey's technology for onward sale to their clients [3]. Managing that internally means taking bets, because a horizontal feature raises value for everyone while a specific workflow is binary, high value if it lands and zero if nobody buys it [3]. His three filters are whether firms and their clients will act as design partners to prove repeatability, whether the technology is there, and appetite, meaning whether the in-house team is genuinely comfortable with AI handling that use case at all, which matters a great deal in compliance and insurance-adjacent work [3].

Behind this sits his read on legal pricing. What clients actually buy is the partner's judgment, "should you buy that company? How could you buy that company? What is the best way to do this negotiation," and the value of a firm is to "do the deal," not to complete each small piece of it [6]. Yet the best partner in the world bills around $3,000 an hour while a junior associate fresh from law school bills $1,100, so the work gets bundled into a pyramid when the thing being paid for sits at the top, which is why he expects pricing mechanisms to change [6]. He has also pushed back on displacement fears by pointing at budget headroom, arguing that legal technology spend moving from around 2.5% to 5% of total industry spending is enormous growth that does not require a massive reduction in the workforce [2], and he has argued that judgment becomes more valuable precisely as routine work is automated [9].

On growth as a pure product story

Asked what took Harvey from $100 million ARR to nearly $300 million in a year, his answer is blunt: "It's 100% product" [1]. The supporting numbers are usage. Token consumption was 1 trillion in January and around 12 or 13 trillion by the time of the conversation; DAU over MAU went from roughly 36% to 51 or 52%; queries per user and hours spent have each been doubling quarter over quarter [1]. He prefers hours to queries as a metric, because a more complex product produces hundred-page outputs, so users may query less while spending more time reviewing and collaborating inside the system [1]. The single largest recent step change was infrastructure, switching the entire stack to another cloud provider, after which usage started doubling again [1][2]. Roughly 2,000 customers, about 42% of them in-house corporates, a segment now growing faster than firms despite having started a year later, with financial services the fastest growing vertical and pharma close behind [1]. That divergence is pushing him to verticalize, since the compliance and legal needs of a bank differ sharply even from private equity [1].

Geography followed customers and regulation rather than a plan. Signing something like Deutsche Telekom created the need for a German office, and because Harvey processes sensitive data, many countries required a local Azure instance, with Australia's restriction on processing financial data offshore as the example; setting up an instance became a reliable leading indicator that an office would follow [1]. The other lens is legal TAM by number of lawyers per country [1]. On capital, he says most of the more than a billion dollars raised has not been spent, and the reason to spend it now is post-training [1]. Legal data for that was unobtainable, since documents from a Blackstone fund formation simply are not online, but the last generation of coding models can generate synthetic documents that lawyers cannot distinguish from the real thing, which has let Harvey build a synthetic data pipeline across every legal use case and begin post-training in earnest [1].

On scaling breaking, again and again

"The problem every single time has been scaling," he says, with a different fix on each occasion [2]. The clearest example was a sev zero in spring of the prior year, when the platform went down for about an hour every other day, caused by an explosion in usage compounded by the launch of Vault, which let users ask fifty questions across 100,000 documents; people uploaded terabytes and there were no caps [2]. The response was to hire a CTO with a deep infrastructure background and principal engineers to match, and by the time of the conversation he expected to be able to support roughly a hundred times current usage within weeks [2]. His general model is that "you go through the process that a company normally would go through in 10 years, like every six months," breaking everything, then arriving at the next doubling prepared [2]. The same pattern has hit product design, roadmapping, sales, go to market and post-sales [2]. What outsiders miss is the timing: "when you're running the company, you see scale breaking six months before it starts breaking," so the founder is having the panic attack about a problem nobody else can see yet [2].

Structurally, he says the roadmap itself has been remarkably consistent; what collapsed was how long things take to build, and what moved fastest was the shift toward a platform underneath the separate pieces [1]. The current organisational version of that is splitting the product into a shared platform plus distinct surfaces and admin features for law firms and for in-house teams, having previously oscillated between quarters focused on user delight and quarters catching up on admin, innovation and customisation features [4]. The middle zone, how legal services are actually delivered and collaborated on across firm and client, is where he expects the most long-term innovation [4].

On hiring, ownership and bias for action

He calls hiring the key unlock, and includes development, promotion, moving people into different roles and, when someone is outscaled, finding a better role for them [1]. His first screen is now singular: "my number one hiring criteria is literally just bias for action," meaning a willingness to act, learn from the action and iterate [5]. The corollary matters as much: "I very much don't penalize mistakes, I penalize in action," and he has had to correct the impression that people were in trouble for a wrong call when the actual problem was taking three months or six to decide, by which point the market and the org had both changed [5]. For senior hires, the test he applies is whether a candidate can map their org at three months, six months and a year, naming every hire, what they do and who might fill the role. "It is insane how many people can't do that" [5]. Leaders who fail to scale are usually the ones who never built leverage around themselves, especially in the glue parts of their function, and much of his own time now goes to identifying and hiring those lieutenants alongside his leadership team [5].

Ownership is his theory of titles and of failure. He prefers "head of" to director because it conveys responsibility for the box rather than partial involvement, and thinks of all titles as ownership, since the instinct everyone has is to squirm out of it [5]. He accepts the plant analogy: two friends asked to water it means it dies from drought or drowning [5]. The same avoidance shows up in roadmaps, where P0s multiply, and someone recently invented a P00 to escape choosing, which he mocked, because three P0s let you diffuse blame when nothing gets used [5]. He admits to making decisions insanely fast, and that his real failure is not deciding who the decision maker is, which leaves six DRIs on a problem and no one deciding [5]. He has massive trust issues and is working on them, notes that his best hires came after being forced to do the role himself for three months, and says that with zero time in a role you will hire the wrong person 100% of the time [5]. His communication has widened from a seven-person leadership Slack channel to a roughly thirty-person group of VPs, heads of and strong ICs, though not the whole company, because raw speculative thinking posted in general chat gets distracting and creates silos [5]. On his own attention, he concedes the standard CEO failure of ignoring what works, and reframes the fix as noticing that a well-run org could be ten times bigger with more resources or a promotion, rather than as remembering to celebrate [5].

Every four months, he says, he hits a mental block where too many things are going wrong at once, which signals a leadership hire, a structural change or something to cut: "you have to just reinvent yourself as a founder every like four months or so" [5]. He describes it as a pressure system that feels relieved for a month and rebuilds over the next three [5]. His own role has swung hard with it: early on his co-founder Gabe ran engineering while he ran everything else and spent most of his time on go to market; in 2025 that became roughly 60% product and 40% GTM; from November it has been about 95% product, engineering and design [2]. He has learned mostly from failure and from watching people he respects, including several he has never met, and names Brian Halagan and Michael Dell among those he has learned from directly [2].

On being doubted, and why it helps

He treats the recurring "Harvey is going to zero" narrative as an asset. A tweet to that effect with roughly a million impressions has appeared every six weeks or so for three years, to the point that he gets nervous when six weeks pass without one [2]. Much of the criticism comes from people who have never seen a demo or used the product, including self-described top haters and people asserting that a general model does the job better [2]. What he values is the compounding effect on the team: the first time it happens the company freaks out and then does well, and by the tenth time the reaction is to prove the doubters wrong. "If you face none of those, you're going to have a team that doesn't have a backbone" [2]. He names this as Harvey's biggest advantage in competing with the model providers, says there is no universe in which the hard times are behind them, and only wishes the labs had come after them earlier, because the threats grow stronger at every stage and the backbone is easier to build early [2].

On the temptation to one-shot the company

The clearest failure he tells against himself came in early 2024, when he and his co-founder were asking whether there was a way to one-shot the company, meaning a strategic play or two that would guarantee success [6]. They flew out to acquire a company with roughly ten times their headcount for close to Harvey's own $700 million valuation, planning to finance it afterwards in the style of a leveraged buyout. They locked the company up, could not quite pull the financing together, and declined to take on the debt required [6]. The week afterwards reset the company: "running a company is just you work 100 hours a week for 10 years and you just slowly build it," and the only route is "consistency and execution at scale" [6]. The next six months went into leadership hires and scaling systems, and he stopped doing every deal himself, which he thinks gave investors more conviction than anything else [6]. Asked how the business would look had the deal closed, he says much worse, because they would have over-indexed on the acquisition instead of building the team and the systems [6][7]. He has since faced similar shortcuts and chosen the harder path of fixing internal systems, and says there have been many more failures like it [6].

That posture is continuous with how he started. What gave a junior associate the confidence to run a company was, in his own one-word answer, "Ignorance" [6]. He had never managed anyone, and his view is that an application layer company forces you to learn in months what would normally take two or three years, because revenue doubling roughly every six months changes what kind of leader and company you need to be [6]. He compares it to learning a language by moving to the country, and says he fails at about 98% of things on the first attempt [6]. The founding itself came from testing chain-of-thought prompting on 100 California landlord-tenant questions pulled from Avvo, showing the answers to three landlord-tenant attorneys with no mention of AI, and finding that on 86 of them all three said they would send the answer to the consumer with no edits. He packaged that up and emailed Sam Altman and OpenAI's then general counsel Jason Kwon to ask whether they knew how good their models were at legal, fully expecting to be told to back off [6][7].

He also had a clear read on why the two standard investor objections were wrong. On hallucination, professional services are already structured as review flows, with a partner decomposing a client question into sub-questions passed down a chain of associates, so "you already have a checking process in in the flow and in the nature of the business" [6]. On lawyers not buying technology, his explanation is that "there hasn't been a technology in the past that impacted the practice of law," which is why legal tech historically addressed the business of law, invoices and timekeeping, rather than the work itself [6]. Once lawyers understood the tool as automating tasks that are not where their value sits, the objection softened [6]. He remains cautious about pace, saying generative AI will fundamentally change how law is practised but much more slowly than the media suggests, which is part of why Harvey invests in integrating with legacy systems and the Microsoft suite that lawyers already live in [4].

On culture, the office and the brand

Harvey has grown to roughly 960 people across 12 offices, with about 350 in San Francisco and 300 in New York and the vast majority of engineering, product and design in San Francisco [1]. Over 200 lawyers work at the company, though only about 25 do what a normal lawyer would do commercially; the rest work on product or go to market, and Harvey uses its own product internally for much of the remainder [1]. The daily ritual he protects is lunch, loud enough that he refuses to book external meetings between noon and 1:30, carried over from the Airbnb days when everyone ate together, and deliberately unsegregated so people meet colleagues across functions rather than clustering by team [1]. He is sceptical of forcing attendance: "I think that if you have to try to force it, you're probably in trouble" [1]. People come in on weekends, which he knows because he loses his badge constantly, is on his nineteenth, and has always managed to find someone in the building to let him in [1]. He works from the couch rather than a room, arriving early to say hello for a couple of hours before meetings [1]. The brand and the office deliberately blend classical references, orators, sculptures, books, with a tech company aesthetic [1]. And an early running joke captures the product obsession: a tiny gavel used to hammer the table whenever a new feature was tested, with the instruction to make it faster [1].

Takeaways

• Value in application layer AI comes from the mess: process data for legal tasks does not exist online, so Harvey hires domain experts to specify the steps and does task-specific post-training rather than dumping legal documents into a model [3].
• Product doctrine is a cycle of expansion and collapse: build specific agentic workflows for high-value tasks, then chain and route them behind one surface so the user never has to find them [3][4].
• Go after the most demanding customers first, because trust in professional services flows downhill from the most prestigious firms to everyone else [3][4].
• Hallucination was never disqualifying in legal, since firms already run a partner-to-associate review chain, and lawyers ignored past technology because it addressed the business of law rather than the practice of it [6].
• Growth is a usage story: tokens went from 1 trillion a month to 12 or 13 trillion, DAU over MAU from about 36% to roughly 52%, and hours spent is a better metric than queries once outputs run to a hundred pages [1].
• The recurring company-killer is scaling, and the discipline is to run a decade of organisational change every six months and break everything on purpose [2].
• Hire for bias for action, penalise inaction rather than mistakes, and require senior candidates to map their org at three, six and twelve months with named roles [5].
• There is no strategic shortcut: the attempted leveraged acquisition of a company ten times Harvey's size at its own $700 million valuation fell through, and the lesson was consistency and execution at scale [6][7].