Winston Weinberg

1. Key Themes

Token Usage as the Truest Growth Metric

Harvey's most striking growth signal isn't ARR — it's raw token consumption, which has grown 12-13x in under a year. This is a proxy for genuine product engagement, not just contract value.

"Our token usage in January was 1 trillion, like for the month of January. And this month it'll probably be like 12 or 13 trillion." 00:00:13

DAU/MAU as the Real Stickiness Signal

Winston explicitly tracks DAU over MAU as his north star for product health, and the improvement from 36% to 51-52% signals that Harvey is becoming a daily workflow tool, not a novelty.

"Our DAU over MAU at the beginning of the year was around like 36%. Right now it's like 51, 52%. Our queries per user have basically been doubling up quarter over quarter. Hours spent is doubling quarter over quarter." 00:00:13

Claude Agents as the Infrastructure Catalyst

The single biggest product lever Winston identifies is migrating the entire infrastructure to Claude agents. This is a non-obvious operational insight — it wasn't a feature, it was a backend architectural swap that unlocked compounding growth.

"This year most recently the main switch we did is just went over to cloud agents. We switched our entire infrastructure to that. And once we did that usage literally just started like doubling quarter over quarter." 00:00:13

Vertical AI Models as the Margin and Moat Strategy

Winston frames the long-term business model around building domain-specific models that outperform frontier models on legal tasks at a fraction of the cost. This is Harvey's answer to both margin compression and lab competition.

"Can you actually create models that do diligence, do change of control review, do contracting, do these different things at the level, if not better than the frontier models. And they cost a hundred times less." 00:29:39

Synthetic Data as the Legal AI Training Breakthrough

Because sensitive legal documents (e.g., Blackstone fund formations) don't exist in public datasets, Harvey has built a proprietary pipeline to generate synthetic legal documents that lawyers cannot distinguish from real ones — unlocking post-training at scale.

"You can actually take sets of documents and create synthetic docs that are so good that the lawyers can't tell the difference between whether they're created by an actual lawyer or they're created by the coding models. And so with that, we've now actually created basically like a pipeline for creating synthetic data sets across like every single legal use case." 00:16:15

The Labs Are the Real Competitor

Winston explicitly names Anthropic (Claude for Legal) and hints at OpenAI (Codex for Legal rumor) as the existential competitive threat — not other legal AI startups like Leiga.

"I think our main competitor is the labs. You saw Claude for Legal. There's a rumor that Codex for Legal, right? I think that the reality is like the labs have an incredible amount of resources. And so it is a race for how quickly can we build the best product that is verticalized." 00:31:29

Reinventing the Company Every Six Months as Survival Mechanism

Winston describes a recurring ~6-month pressure cycle where things "break," requiring three major structural changes each time. He treats this as a feature, not a crisis — and uses it as a filter for who can scale with the company.

"Every six months, I'd say I start to get like this weird feeling of like things are just breaking. And then usually what has happened is every time that has happened, there's been like three big changes that I need to make and I realize I need to make them. And then I make them and then the pressure flows off." 00:17:57

The ROI-per-Token Problem Is the Defining Business Challenge of the AI Era

Winston draws a direct analogy between the law firm billable hour (granular ROI justification for every 6-minute increment) and what every enterprise AI buyer will soon demand: proof of return on every token spent.

"I just spent a billion dollars on tokens. Where's my ROI? Right? And I actually think that like, there aren't enough companies that I know of that are starting to think about, how do I actually show ROI in every single vertical use of these things." 00:33:52

Office Expansion Driven by Data Sovereignty Requirements

Harvey's geographic footprint wasn't primarily sales-led — it was forced by country-specific data localization laws, which required standing up Azure instances per jurisdiction before customers could be onboarded.

"In Australia, you can't process financial data outside of the country. And so we would set up these offices and then we'd set up like Azure instances too. It was almost like we set up an Azure instance and then that would be like a pretty good indicator that we'd have to set up an office pretty soon afterwards." 00:04:44

Acqui-Hire Strategy: Talent Over Technology

Winston is explicit that Harvey's M&A thesis deliberately avoids buying legal tech — instead targeting high-quality teams from outside the vertical. Legacy technology is considered a liability in the current build-speed environment.

"The aqui-hires that we have done, they actually haven't been in the legal AI space or legal tech, they've been outside of it, but they're really, really good teams that could work on a problem that we have." 00:24:17

2. Contrarian Perspectives

Vertical Models Will Beat Frontier Models on Price-Performance — and That Gap Is the Entire Market

Most people assume frontier models (GPT-5, Claude 4, etc.) will eventually become cheap enough to use universally. Winston argues the opposite: that there will always be a massive economic space between frontier and commodity models, and the entire vertical AI industry lives in that gap — including scenarios where GPT-10 costs more than a human lawyer.

"There's a world in which GPT-10 is more expensive than a lawyer. That's a very possible world. And so where we have is basically you can think of like the frontier models here, commoditized models here. I think a lot of the economy is actually in between these things." 00:30:43

The Labs Are Every Company's Real Competitor — Not Other Startups

Conventional startup analysis focuses on horizontal competitors in the same vertical. Winston flips this: the true competitive threat is the foundation model labs entering every successful vertical directly. The better any vertical AI does, the more resources labs will redirect toward it.

"I really do think at the end of the day, it is a race against the labs. And I think every single company on earth is competing against them... The better we do, the better other companies do, the more the labs will be like, oh, okay, we're going to put more resources into that." 00:31:54

Existing AI Benchmarks Are Almost Entirely Useless for Vertical Applications

Winston argues that outside of coding, no vertical has a benchmark worth trusting. Legal benchmarks in particular test for bar exam trivia rather than real professional tasks — meaning most published "accuracy" claims for legal AI are measuring the wrong thing entirely.

"If you look at like the benchmarks that have existed for legal for a long time, I mean, half of them are like, can it pass the bar? Multiple choice questions on community property law. And I think that we haven't actually until now had a very good set of data that does a legal task from end to end. And we're missing this in most verticals other than coding." 00:34:30

Don't Buy Legacy Technology in M&A — Even in Your Own Vertical

The standard M&A playbook for category leaders is to consolidate competitors. Winston explicitly rejects this, arguing that build speed has made existing technology nearly worthless as an acquisition asset — even in legal tech — and that talent is the only real thing worth buying.

"I do not believe that it is a good idea right now to go around and buy legacy technology. I think it is a much better idea to buy really, really good teams regardless of if they worked in your space." 00:23:49

Making Decisions at 51% Certainty Is the Required Operating Mode

Most enterprise operators pride themselves on data-driven, high-confidence decision-making. Winston argues this is now a liability — the pace of AI development requires treating most decisions as reversible and acting on thin majorities of evidence, accepting a high error rate as normal.

"You're probably going to have to make a bunch of decisions at like 51% certainty, which means you're being wrong. And can you deal with being wrong and then quickly pivoting? And I think you have to do that way more than you used to in the past." 00:27:18

3. Companies Identified

Harvey AI

Legal AI platform. The subject company — $300M ARR, 960 employees, 12 offices, 2,000 customers, ~$11B valuation, processing 12-13 trillion tokens per month. DAU/MAU at 51-52%.

"We're around 2,000 customers, somewhere around 300 million ARR. We're growing really fast." 00:13:37

Anthropic

Foundation model lab. Winston credits switching to Claude agents as the infrastructure catalyst behind Harvey's doubling of usage quarter over quarter. Also released "Claude for Legal," which Winston names as a direct competitive threat.

"The main switch we did is just went over to cloud agents. We switched our entire infrastructure to that. And once we did that usage literally just started like doubling quarter over quarter." 00:00:13

Deutsche Telekom

German telecommunications giant. Cited as an example customer whose signing triggered Harvey opening a Germany office and standing up a local Azure instance.

"We'll sign like Deutsche Telekom and it's like, oh wow, we need an office in Germany." 00:04:21

Blackstone

Global alternative asset manager. Used as the canonical example of why legal training data doesn't exist publicly — sensitive fund formation documents never leave the firm.

"If you went online and you're like, I want to go find a bunch of documents related to like a random fund formation by Blackstone — they don't exist. You'd have to go to Blackstone for those documents." 00:15:45

Microsoft Azure

Cloud infrastructure provider. Harvey uses Azure for jurisdiction-specific data sovereignty compliance, standing up separate instances per country before entering new markets.

"We would set up these offices and then we'd set up like Azure instances too." 00:04:44

Applied Intuition

Autonomous vehicle software company. Mentioned favorably as another company with a notably strong in-office culture.

"The only other company that was applied intuition. They were packed." 00:09:40

Leiga (Lagora)

Legal AI competitor. Mentioned as a fast-moving competitor claiming European dominance, which Winston contests, citing Harvey's over 70% win rate in Europe.

"I would say in Europe, I think our win rate is like over 70%. So I don't think it's as much dominant there." 00:31:29

Uber

Used as a contemporary case study for the intelligence allocation problem — specifically around whether frontier AI is required for every task or whether cheaper models suffice.

"The best example of this is like, what happened with Uber recently, right? And it's like, we're going to get to this point." 00:32:55

Brex

Sponsor. Financial platform for startups — combines checking, treasury, and FDIC protection. Trusted by one in three venture-backed US startups.

"The financial stack trusted by more than 30,000 companies, including one in three venture-backed startups in the US." 00:11:25

Scale AI

AI data and infrastructure company. Named as a Brex customer.

"Companies like Scale AI, DoorDash, Service Titan, HIMSS, Anthropic, Flexport, Robinhood, and Plaid trust and use Brex." 00:11:52

Cholita Linda

Local San Francisco restaurant. Winston's habitual lunch and dinner order, placed 467 times on DoorDash.

"It's the place called Cholita Linda and it's around here." 00:35:57

4. People Identified

Pat Grady

Partner at Sequoia Capital (Harvey investor). Praised Winston specifically for Harvey's ability to repeatedly reinvent itself — cited as the defining quality of the company.

"I reached out to Pat Grady and he said, the one thing that he emphasized was that you've been able to reinvent the company over and over again." 00:17:35

Co-founder and CEO of Harvey AI. Former lawyer turned AI entrepreneur. Built Harvey from an Airbnb to a $11B company processing 13 trillion tokens/month in under four years. Hands-on on product, office design, hiring philosophy, and competitive strategy.

"It'll be four years in August. Close, it's like 960 something [employees]." 00:03:32

5. Operating Insights

The Calendar Audit as the First Diagnostic for Scaling Leaders

When a leader starts breaking under scale pressure, Winston's first intervention isn't a reorganization or coaching plan — it's a granular audit of their calendar, forcing them to justify every item against a single stated priority for the week.

"When I see somebody massive starting to really break, the first thing I do is I'll go and we'll do like a calendar audit. You go in and you do a calendar audit and you come up with really good ways to be like, do you have to do any of these things? Tell me what is the main priority this week. Are any of these related to that priority?" 00:19:15

The One-Paragraph Rule for Separating Real Priorities from Noise

Winston uses a simple forcing function with his chief of staff: before any non-routine commitment is accepted, he must write a paragraph justifying it — without AI assistance. If it's hard to write, it shouldn't be on the calendar.

"If you sit down and you can write an entire paragraph about why you're going to do something, you definitely are going to do it. If what happens is you go, oh, it's so annoying that I have to write this down — you probably don't need it." 00:20:30

Loud Communal Lunch as a Culture Preservation Tool at Scale

Harvey deliberately protects a daily shared lunch culture — now at nearly 1,000 employees — where all functions mix without seating by team. Winston treats this as infrastructure for organic cross-functional trust, and refuses to schedule external meetings during it.

"I try to make it so you can't be like, oh, that's the team that works on this — it's like everyone is all mixed together. I think people like coming into the office because we have a very culture of everyone working together and people are really good friends that are across functions." 00:08:07

Hiring Lawyers Who Loved Law But Hated Big Law — Then Redeploying Them as PMs

Harvey's lawyer hiring strategy targeted people who were frustrated with the BigLaw structure (not the practice of law), then created career paths that allowed them to transition fully into product roles. This produced a deeply domain-literate product organization without the cultural friction of traditional legal hires.

"We have so many lawyers that are now just full-time PMs, like literally they just transferred. You could have a legal background and then you end up doing something else. You don't actually have to be practicing law here." 00:21:29

Azure Instance Deployment as a Leading Indicator for Office Expansion

Harvey discovered that data sovereignty requirements forced Azure instance setup before customer onboarding in new countries — making Azure deployment a reliable forward indicator of where office infrastructure would be needed. This turned compliance into a predictable expansion sequencing tool.

"We set up an Azure instance and then that would be like a pretty good indicator that we'd have to set up an office pretty soon afterwards just because of customer demand." 00:04:44

6. Overlooked Insights

The ROI-per-Token Problem Is Bigger Than Legal — It's the Defining Business Model Question for All of Enterprise AI

Winston mentions this almost as an aside, but it is arguably the most important structural observation in the entire conversation. Every enterprise buying AI is about to face the same accountability problem that law firms solved decades ago with the billable hour: granular justification of every unit of compute spend. The companies that build vertical ROI attribution layers — showing exactly what each token delivered — will have a structural advantage in renewals and expansion that has nothing to do with model quality.

"I just spent a billion dollars on tokens. Where's my ROI? Right? And I actually think that like, there aren't enough companies that I know of that are starting to think about, how do I actually show ROI in every single vertical use of these things. And I think that vertical companies are going to have a huge advantage here, where you can start to get to the point where you basically can show every single token and what the ROI was of that token for your particular task in vertical." 00:33:52

Good Legal AI Benchmarks Don't Exist Yet — Creating Them Is a Moat

Winston flags in passing that outside of coding, no vertical has a credible end-to-end benchmark — and that Harvey is apparently building one for legal. This is non-obvious because whoever owns the benchmark owns the narrative about what "good" means in that vertical. If Harvey publishes a rigorous legal benchmark, it becomes the de facto standard, giving them the ability to define the competitive evaluation criteria for every competitor and every customer procurement process.

"We haven't actually until now had a very good set of data that does a legal task from end to end. And we're missing this in most verticals other than coding. Basically coding is the only one that has a good saturated benchmark." 00:34:56