# Harvey uses synthetic legal data to train models

Markdown copy of https://realalpha.app/s/harvey-uses-synthetic-legal-data-to-train-models/ on Real Alpha, the best minutes from long interviews with the people building AI. Cite and link the page, not this copy.

- Speaker: [Winston Weinberg](https://realalpha.app/builders/winston-weinberg/) · Winston Weinberg is cofounder and CEO of Harvey. · Co-founded and leads Harvey, after working as an associate in securities and antitrust litigation at O'Melveny.
- Host: [Molly O'Shea](https://realalpha.app/people/molly-oshea/) · Founder, Sourcery · Ex-Upfront Ventures investor; has interviewed Alex Karp and Palmer Luckey.
- Show: Sourcery
- Episode: [Inside Harvey AI: CEO Winston Weinberg on How a 4-Year-Old Company Is Taking on the Foundation Labs](https://realalpha.app/v/ekXJCDNEmOM/), Jun 16, 2026
- Moment: 13:00 to 17:39 of the episode (5 min), playing on https://realalpha.app/s/harvey-uses-synthetic-legal-data-to-train-models/; the episode from there: https://realalpha.app/v/ekXJCDNEmOM/?t=780

## Quote

> “we've now actually created basically like a **pipeline for creating synthetic data** sets across like every single legal use case”
> — [Winston Weinberg](https://realalpha.app/builders/winston-weinberg/) at [16:29](https://realalpha.app/v/ekXJCDNEmOM/?t=989) · Sourcery

## Takeaways

1. ARR rose from $100M to about $300M in under a year.
2. Monthly token use went from 1T in January to a projected 12–13T.
3. Synthetic datasets help train models for legal tasks with scarce source data.

## About this segment

Winston Weinberg ties Harvey’s growth to rising product use, then explains how it builds synthetic legal datasets for model training. The pipeline addresses a shortage of real legal documents for training.

From “Inside Harvey AI: CEO Winston Weinberg on How a 4-Year-Old Company Is Taking on the Foundation Labs”, Jun 16, 2026. [Watch the full episode](https://realalpha.app/v/ekXJCDNEmOM/)

## Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

[13:05](https://realalpha.app/v/ekXJCDNEmOM/?t=785) Okay. So where is Harvey at today? Give me the full stats. Yeah. So we're about four years old, August. That'll be the birthday. I think it is birthday. What's the day? August 4th. It's a Leo. Holy crap. Wow. Actually, I didn't think about it that way. Yeah. Huh. Well, that puts like everything into perspective for me i gotta like take 20 minutes to have like deep thoughts on that i'm a capricorn so i don't know if that works is that a problem i have no idea i don't either i have no idea leo's like a lion right yeah fire sign fire um so yeah about four years old um and uh we're around 900 uh folks uh 950 people 960 something like that um we're at around 2 000 customers um somewhere around 300 million ARR um so we're going really fast um I think like it's it's mostly just a crazy experience like how much the product has grown um I think that in the

[14:01](https://realalpha.app/v/ekXJCDNEmOM/?t=841) beginning like we had this product roadmap and our product roadmap has actually like really been pretty consistent the main thing that's collapsed is just like how long it takes to build these things um and so I think the the thing that's actually moved the fastest is how much we are starting to basically build a platform for all of these different pieces of the product so you were at 100 million in ARR last August and now you're nearing 300 million what's been driving that uh it's 100 product I mean like I'll give you some metrics um 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 oh my god yeah so the usage is getting pretty insane 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

[14:52](https://realalpha.app/v/ekXJCDNEmOM/?t=892) doubling up quarter over quarter hours spent is doubling quarter over quarter I think that's better than queries the problem with queries is as you build a more complex product you're gonna have these outputs that are like crazy long and so you might have people like using this like querying less but the output is like a hundred page document and so they're like in your system reviewing it and collaborating on it um but um like everything is just product improvements i think like last year last year we had some moments where we just kind of had to like rebuild a lot of our product um and then this year most recently the main switch we did is just one order cloud agents like we switched our entire infrastructure to that and once we did that usage literally just started like doubling quarter over quarter and you've raised over a billion dollars

[15:38](https://realalpha.app/v/ekXJCDNEmOM/?t=938) now so have you used most of that capital whereas are you using it on tokens like what is yeah yeah now we're using the office now we're using on tokens um no we actually we haven't used a lot of that money um i think like the the interesting thing that we always wanted to do was actually do a lot of post training on the models and there's a couple problems in legal that makes us like really hard One is the data isn't available. So like if you went online, you're like, I want to go find a bunch of documents that are related to like a random fund formation by Blackstone. They don't, they don't exist. Like you'd have to go to Blackstone for those documents. Right.

[16:16](https://realalpha.app/v/ekXJCDNEmOM/?t=976) But the thing that happened with like the last generation of coding models is 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 you know 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 and because we have that now now you can start actually plus training models and that's me expensive Wow and so what are the main categories of your customers now I think you said it's 42% is in-house corporates yeah 42% is in-house corporates it's growing faster than the law firms technically um because i think we're we're at like almost 70 of the is it easier for

[17:01](https://realalpha.app/v/ekXJCDNEmOM/?t=1021) them to adopt is that why no they're just slower to adopt yeah so the corporates adopted like they started like a year after the law firms um and so and then our fastest growing of like the of the verticals in general is uh financial services is number one for sure so like banks private equity asset management um and then pharma is actually growing pretty fast too uh and what we're starting to do is actually like we're going to verticalize our product too where we have to actually like have different parts of harvey that are different for each vertical right because like the compliance and legal needs of a bank are very very different than even private equity so i want to go into