Winston Weinberg on Building Harvey, Enterprise AI and the Art of Deals

Harry Stebbings welcomes Winston Weinberg, Harvey's CEO and co-founder. Weinberg says OpenAI was the company's first investor, followed by angel investors Sarah Guo and Elad Gil. His comments throughout reflect the perspective of a founder whose company sells AI tools to legal and other professional-services customers.

Weinberg describes taking up a daily mile run when startup stress mounted. He worked on reducing his time and found that a physically demanding start to the day helped him manage stress and make decisions. That is his personal routine, not a universal prescription.

Getting up early has also given him uninterrupted time to think about product before messages begin arriving. With customers across time zones, he tries to keep an East Coast schedule, even rising around 4 a.m. in San Francisco.

His less helpful habit is checking every Slack channel repeatedly. Knowing everything helped him decide quickly in the early company, he says, but as Harvey grows it distracts from the highest-priority issues and makes delegation harder.

Weinberg recalls initially mistaking other founders' practice of dropping into meetings for a lack of work. He later saw a different lesson: they had built organizations that could operate without constant heroics, leaving founders free to inspect the machine and concentrate on product.

He describes a cycle: first find product-market fit, then build a company structure suited to the market, then return to product-market fit as the opportunity changes. He says Harvey has moved through those stages and he is again focused on where the product should go over the next six months to a year.

Stebbings cites approximately $190 million in annual recurring revenue and an $8 billion valuation. Weinberg says the team should benchmark itself against the wider AI market and not mistake market demand for execution alone. He expects rapid change and says he wants more growth than the host's illustrative projections; those figures are targets and estimates, not realized revenue.

Asked which fundraising round felt uncomfortable, Weinberg recalls the Series C at roughly $1.5 billion, when revenue was lower. He says he has often chosen investors he trusts over seeking the highest possible valuation, though that is his account of the company's decisions.

His preferred fundraising approach starts months before a round. A small early investment and information rights let a prospective partner watch the business and compare promised milestones with results. Consistent delivery can build enough trust to shorten a later fundraising process, but he acknowledges that a targeted process may sacrifice price competition.

Investors have helped him see when senior leaders were needed, Weinberg says, and he believes he sometimes hired too late. He is less convinced by their recommendations about whom to hire: boardroom polish and a strong résumé do not necessarily reveal how someone actually runs a team.

He disputes the idea that a famous venture firm can simply make a winner. Money cannot compensate for bad product decisions, he argues, and many legal customers do not recognize Silicon Valley investor brands. In Europe, a firm such as EQT may be more familiar to lawyers than a technology venture brand.

Investor names may nevertheless help recruiting, Weinberg concedes, because candidates use prestige as a shortcut for judging companies. He worries that joining mainly for the investors can signal less commitment to the mission, an assessment he offers as a personal hiring judgment rather than a rule about every candidate.

Stebbings asks whether commitment matters if a sales leader can reliably hit a number. Weinberg says the experience inside a fast-growing startup is often messy even when it looks successful from outside: problems, perceived threats and morale swings are constant. In his view, a connection to the mission helps people persist through that uncertainty.

The threat he thinks about most is product speed. As OpenAI, Anthropic and other model providers improve, an application must offer enough value beyond a generic enterprise model license to remain useful. He frames the challenge as creating a product advantage faster than the underlying platforms advance.

Harvey routes requests among models according to the use case, and Weinberg says traffic to Claude Opus 4.5 rose substantially. OpenAI is an investor, but he says Harvey is not required by its agreement to use OpenAI models and was not routing a majority of traffic to Anthropic at the time of the conversation.

Weinberg distinguishes a possible plateau in some consumer-facing model uses from enterprise work. He believes everyday consumers often need better context and connections to their applications more than additional reasoning, while coding and enterprise capabilities can continue improving quickly. Both the distinction and the growth outlook are his predictions.

Asked about hypothetical valuations for Anthropic and OpenAI, he is bullish on both but qualifies how much he would pay for OpenAI. He sees its consumer brand as powerful and expects multiple enterprise model suppliers; OpenAI is also a Harvey investor, which is relevant to his perspective.

His central argument is that capability can outstrip adoption. Even if model development stopped, he expects businesses and consumers to spend years integrating existing tools into daily work. He estimates that substantial enterprise productivity gains could take three to five years, an outlook rather than a measured result.

A typical enterprise workflow draws on many separate systems, making end-to-end automation difficult. Weinberg says Harvey has seen corporate departments beyond legal, including compliance and procurement, begin using its tools even before the company built features specifically for them.

Harvey's shared spaces were intended to let corporate legal teams and law firms collaborate, and he says internal departments also use them together. When Stebbings relays competitor Legora's claim that the idea was copied, Weinberg denies it and says Harvey had spent months on security and permissions before releasing the interface. The conversation does not independently resolve that dispute.

Weinberg says he respects Legora and credits its early work in Europe. He wishes Harvey had invested in the region sooner, but says a very small team had limited bandwidth: its first large customer rollout involved thousands of users when the startup had only four people.

A polished demonstration is not enough for enterprise deployment, he says. Harvey added tens of thousands of users in late 2023 and discovered that its infrastructure was not ready, slowing its shipping in early 2024. He says the company subsequently invested heavily in senior infrastructure engineers.

He argues that investors should pay attention to gross revenue retention, not just new annual recurring revenue. If a company signs customers rapidly but cannot implement and support what it sold, customers may leave. His concern is a forecast about the sector, not evidence that any named competitor has failed.

Enterprise growth changes the organization as well as the product. Weinberg points to Microsoft's post-sales investment as an example and argues that retaining customers can create opportunities for much larger relationships over time. His example of a customer growing from $1 million to $100 million is hypothetical.

Asked whether customers should be difficult to leave, he argues that a better relationship comes from producing enough value that the customer wants to stay. He describes law firms using customized tools to compete for work and corporate departments saving time, while acknowledging the industry's existing hourly billing model.

Weinberg expects some law firms to experiment with fixed fees and says Harvey customers may accept more consumption-based pricing. He presents those as expectations and company observations; the interview does not establish that every customer prefers those models or independently verify the return on investment he describes.

Some customers, he says, fund Harvey from professional-services budgets instead of technology budgets. He estimates Harvey's revenue mix at about 60% law firms and 40% in-house corporate teams at the time, and expects a similar split in five years. The future mix is a prediction, and the current figure is his approximate company disclosure.

On the future of law-firm employment, Weinberg expects new demand for advice to offset some work that clients no longer want to buy, such as routine document markup. He suggests AI risk, cross-border regulation and expanding product lines could create new assignments. This is his forecast, not evidence that junior jobs are protected.

He also expects AI to reshape the wider economy while acknowledging that markets may experience short downturns. His view that professional services will grow alongside economic activity is an argument about future demand, and the effect on individual workers or firms remains uncertain.

For European expansion, Weinberg recommends working with local teams and adapting products to each market rather than assuming a San Francisco playbook transfers unchanged. He says hiring timelines can be longer because of notice periods and garden leave, so office openings require more advance planning.

Asked about stereotypes of European work habits, he says they do not match his experience with lawyers, while noting that his sample is concentrated in an unusually demanding profession. He looks for ownership in hiring: the ability to identify a problem, accept responsibility and be candid about mistakes.

As Harvey grows, he can no longer personally see every source of trouble. When several people point to one another, he says, strong operators stand out by acknowledging what they controlled instead of offering a polished weakness that is really a boast.

Returning to his Slack habit, Weinberg says the deeper problem is difficulty trusting others to handle issues. A founder who cannot rely on other leaders cannot scale effectively, he argues. He links his own independence partly to his upbringing while recognizing he cannot fully separate nature from experience.

He compares leadership to a sports team: people should care about their own contribution because it helps the team win, not simply because they want the highest individual score. He and Stebbings criticize résumé and company-logo chasing, though those generalizations are their personal assessments of technology hiring.

Weinberg recalls a prediction after ChatGPT launched that investors would fall back on prestigious résumés when they did not understand the field. He thinks social signals can become substitutes for judgment under uncertainty, particularly in a fast-moving AI market.

He estimates that truly great AI researchers number in the hundreds and says a résumé alone cannot identify them. His suggestion is to ask respected researchers whom they regard as excellent. That estimate and his characterization of the research community are his impressions, not a measured census.

Asked about researcher departures, he says talented people may leave when a lab changes direction and the problems no longer match what they joined to work on. He explicitly says he is discussing labs generally, not claiming to know the circumstances at Thinking Machines.

Application companies may initially look like wrappers around a model, Weinberg says, but differentiation can come from conventional software infrastructure and custom work using customers' data. As such products mature, he expects demand for specialized AI talent to return alongside ordinary software engineering.

On deal-making, his first rule is to listen. Talking the most can feel like progress, he says, but it does not prove that a negotiation is advancing. Understanding what individuals and groups actually want is more valuable than trying to dominate the conversation.

His second rule is to recognize when not to negotiate every term. If a particular asset or relationship is unusually valuable to the business, he may concede financial terms to secure it. The judgment depends on understanding the value and the trade-off; it is not a claim that every concession produces a good deal.

Weinberg describes deals as interconnected ropes: closing one can relieve pressure and make another possible. He points to Microsoft's partnership strategy as an example of a company benefiting from a broad ecosystem, and Stebbings applies the same logic to avoiding unnecessary bargaining over a valued hire's pay.

He has misread candidates by mistaking weak communication skills for an inability to grow, only to find that those skills could be learned. He also admits falling for the résumé trap. Both examples reinforce his distinction between surface presentation and the underlying ability to contribute.

In 2022, Weinberg and his co-founder cold-emailed Sam Altman and Jason Kwon after testing a legal-answer prototype using questions from an online forum. He describes asking landlord-tenant lawyers to review answers, but the automatic transcript leaves the outcome of the 86-of-100 figure incomplete, so it cannot establish an accuracy rate.

The team first spoke with Kwon and then pitched OpenAI leadership on July 4, 2022. OpenAI became the seed investor. Weinberg recalls an approximate $4 million pre-investment valuation but says he does not remember the precise terms.

At the Series A, he says, the founders met roughly ten venture firms in two days and he knew little about their reputations. His co-founder came from technology, but Weinberg judged the meetings largely on the interaction itself; he remembers approximately half producing term sheets and acknowledges the estimate is uncertain.

Weinberg says a lesson of the past year is that some company-building fundamentals remain unchanged even in AI. For example, he had initially failed to connect a new revenue target to the number of salespeople needed, their quotas and the time required to recruit and ramp them.

He compares his investors' styles: he sees Andreessen Horowitz as larger and louder, and Sequoia as operating with a smaller partnership. Both firms have invested in Harvey, so his account is that of someone with continuing relationships with them.

He keeps an operating document with short reminders associated with people he admires. He attributes to Keith Rabois the idea of deliberately taking on difficult, stressful work and says it helps him; this is his personal management philosophy, not a general health recommendation.

From Pat Grady, he takes the idea of persistent application of effort. From Brian Halligan, he learned to say no. He says product planning should involve declining some attractive ideas so a team can focus on the few it can execute well.

For the coming year, Weinberg wants Harvey to move from a useful productivity application toward infrastructure that legal professionals rely on daily. He says the company has built multiple product lines but has not yet joined them into a coherent platform.

He reports a 74% daily-to-monthly active-user ratio among people using at least four product lines, while acknowledging that this group is still a small share of users. That selective cohort figure is not the engagement rate for all Harvey users. His aim is to integrate the tools and expand sustained use across the broader customer base.