Harvey AI Speaks! - by Cyrus Johnson "AI Counsel"
On February 14, a new episode of the "No Priors" Podcast featuring Elad Gill and Sarah Guo was released, with special guest Winston Weinberg, the CEO of Harvey. This podcast provided a rare glimpse into the typically reserved Harvey, making it a must-listen for those interested in AI developments in the legal industry.
AI COUNSEL NEWS + PODCAST is a reader-supported publication. To stay updated on new posts and support the work, you can become a free or paid subscriber.
Ai Counsel Summary
The podcast discussed Harvey's journey in creating an AI platform for legal and professional services. Founded in 2022, Harvey was inspired by successful experiments with GPT-3 in legal workflows. With over $500 million raised and a client base of more than 250, the company emphasizes the importance of building trust through partnerships with prestigious firms and data providers.
The product strategy at Harvey focuses on a platform that combines productivity tools with streamlined workflows. By creating reusable "AI patterns" for complex operations and simplifying the user interface, Harvey is able to address both general productivity needs and specialized tasks with high-quality outputs.
Team building is a key aspect at Harvey, with senior lawyers serving as design partners and evaluators. The company places a strong emphasis on hiring individuals with adaptability and agency, recognizing the evolving landscape of AI. Harvey fosters a culture of innovation and resilience, with leaders who take ownership and are committed to continuous improvement.
The podcast also touched on the potential impact of AI on the legal and professional services industries, predicting significant changes such as a shift towards fixed-fee models and collaborative efforts between law firms and technology companies. Harvey is also looking to expand into areas like tax and audit services.
Transcript
# No Priors Ep. 101 | With Harvey CEO and Co-Founder Winston Weinberg
00:00:03.070 [Applause]
00:00:05.879 hi listeners and welcome back to no 00:00:07.439 priors today I'm here with Winston 00:00:09.400 Weinberg the co-founder and CEO of 00:00:11.679 Harvey which is building domain specific 00:00:14.280 AI for law Professional Services and the 00:00:17.279 Fortune 500 they've now raised more than 00:00:19.920 $500 million from investors such as open 00:00:23.080 aai seoa finer Perkins GV lill and me 00:00:27.720 we're going to talk about how to do end 00:00:30.400 to-end workflows how to serve 00:00:32.238 conservative users impostor syndrome 00:00:35.120 keeping Pace with the blitz of the AI 00:00:37.559 Eco system and also what lawyers will do 00:00:40.320 5 years from now Winston thanks for 00:00:42.399 doing this yeah of course it has been 00:00:44.440 like a wild two and a half years for you 00:00:46.559 and Gabe and Harvey um when you started 00:00:49.559 the company in August of 2022 or at 00:00:52.320 least when I met you guys for the seed 00:00:53.960 yeah we started a little bit earlier but 00:00:55.160 about then yeah uh what was the moment 00:00:57.239 of inspiration yeah so G and I actually 00:01:00.800 had met a couple years before and I 00:01:03.640 definitely didn't know anything about 00:01:04.680 the startup world and didn't have a plan 00:01:06.479 of of doing a startup and what had 00:01:09.000 happened was he showed me gbd3 which at 00:01:11.840 the time was you know public and and I 00:01:14.119 was first of all just incredibly 00:01:15.960 surprised that no one was talking about 00:01:17.400 gbd3 and no one was using it in any way 00:01:19.400 shape or form um and he showed me that 00:01:22.840 and I showed him kind of my legal 00:01:25.000 workflows and we started the the kind of 00:01:28.000 aha moment was we we went on uh r/ legal 00:01:31.880 advice which is basically you know a 00:01:34.079 subreddit where people ask a bunch of 00:01:36.360 legal questions and almost every single 00:01:38.439 answer is so who do I sue um almost 00:01:40.840 every single time and we took about a 00:01:43.119 hundred landlord tenant questions and we 00:01:46.320 came up with kind of some Chain of 00:01:47.840 Thought prompts and this is before you 00:01:49.640 know anyone was talking about Chain of 00:01:50.920 Thought or anything like that and we 00:01:53.119 applied it to those landlord tenant 00:01:54.840 questions and we gave it to three 00:01:56.399 landlord tenant attorneys and we just 00:01:58.360 said nothing about AI we just said here 00:02:00.799 is a question that a potential client 00:02:02.680 asked and here is an answer uh would you 00:02:05.520 send this answer without any edits to 00:02:07.680 that client would you be fine with that 00:02:09.080 you know is that ethical is it a a good 00:02:11.000 enough um answer to to send and 86 out 00:02:14.560 of 100 was yes um and actually we cold 00:02:18.040 emailed the general Council of open aai 00:02:20.480 and we sent him these results and his 00:02:22.840 response basically was oh I had no idea 00:02:24.879 the models were this good at legal um 00:02:27.200 and we we met with the the SE Suite of 00:02:28.879 open AI a couple weeks after and and the 00:02:31.640 view was just is going to be good enough 00:02:34.360 um we should build a company around it 00:02:36.360 yeah for what domains I mean I think 00:02:38.360 what happened or the reason we were so 00:02:41.000 confident about this was the models and 00:02:44.159 even with gbd3 you could get it to do a 00:02:47.159 lot of tasks you just had to really 00:02:49.360 brute force it right like you had to 00:02:51.080 Brute Force the amount of context 00:02:52.920 telling it which steps to take Etc and 00:02:55.560 the idea was over time this is just 00:02:57.560 going to get better right they are going 00:02:59.159 to be either the models themselves are 00:03:01.159 going to get better or we're going to be 00:03:02.800 we're going to get better over time at 00:03:04.799 figuring out how to provide them the 00:03:06.120 correct context how how to improve them 00:03:08.400 how to evaluate the results Etc um and 00:03:11.560 even just by playing with it for a 00:03:13.319 decent amount of time you could get that 00:03:15.159 sense you are obviously not focused just 00:03:18.040 on property law now like how do you 00:03:20.080 think about the mission or scope of 00:03:21.879 Harvey today yeah um so mostly we're 00:03:25.720 we're de developing it for legal overall 00:03:28.319 but I would say that what we're building 00:03:30.879 is the AI platform for legal and 00:03:32.640 Professional Services right and if that 00:03:35.040 sounds vague or it sounds like there 00:03:37.040 aren't incredibly defined use cases for 00:03:40.000 you know the small areas that we're 00:03:41.400 building that's on purpose like the 00:03:43.280 reality is if you are using these tools 00:03:46.560 and you don't think that you can take 00:03:48.640 basically Ai and apply to X industry and 00:03:51.519 transform the entire industry I don't 00:03:54.040 think you're thinking ambitiously enough 00:03:56.319 right and I think it's really hard 00:03:58.280 because the you know these models can't 00:03:59.519 just onot all of these really complex 00:04:01.879 legal tasks or or in these other domains 00:04:04.040 like tax and Prof other Professional 00:04:06.360 Services and so what you have to do is 00:04:08.599 you have to build a platform that is 00:04:10.680 kind of constantly expanding and 00:04:12.480 constantly collapsing and so what I mean 00:04:14.400 by that is you need to build specific 00:04:17.079 features and maybe agentic workflows Etc 00:04:20.000 that can do parts of a task and then you 00:04:22.160 need to combine them all together so the 00:04:24.040 UI is simple and you don't have this 00:04:25.800 like tentacle monster of a platform when 00:04:28.360 did you realize that that was the like I 00:04:30.080 think it's a really elegant framing of 00:04:32.199 sort of product strategy for the company 00:04:34.199 yeah um uh first implied in it is just 00:04:37.120 like this you know and this has been 00:04:38.639 true for you guys from the very 00:04:39.680 beginning I I did think you were a 00:04:40.759 little bit crazy when you described the 00:04:42.360 level of sophistication of tasks end to 00:04:45.120 end that Harvey would be able to I still 00:04:46.680 feel crazy uh but you appear to be more 00:04:49.919 you know like you don't look as crazy 00:04:51.160 when you're right fair right um but I 00:04:53.680 remember like a year and a half two 00:04:55.240 years ago um uh like it was very 00:04:59.240 distinct to me how much you guys 00:05:00.759 believed in capability Improvement yeah 00:05:02.960 like where does that come from and then 00:05:04.199 also how do you think about it 00:05:05.320 internally in the company yeah so where 00:05:07.800 does it come from using the tools I mean 00:05:10.560 I think I have been blown away by the 00:05:13.440 amount of kind of seite and Etc that 00:05:16.600 haven't actually like spent that much 00:05:18.240 time using Ai and the the big moment for 00:05:21.720 me was just the jump between gbd3 and 00:05:24.319 gbd4 like when we got access to gbd4 I 00:05:27.880 went into my room I think for 24 hours 00:05:30.600 straight and try to do every single 00:05:32.800 thing that I couldn't do with gbd3 or I 00:05:35.400 had to very much brute force and like 00:05:37.360 wasn't getting any efficiency gains from 00:05:39.600 and try to do it with with gbd4 it 00:05:41.600 didn't do all of them but the 00:05:43.080 Improvement was crazy right and no one 00:05:46.039 was talking about it and I I found that 00:05:47.840 something that also was very weird is I 00:05:50.000 would show gbd3 and and gbd4 to a bunch 00:05:52.560 of my friends and they would try to get 00:05:55.199 it to do one thing it didn't do the 00:05:56.759 thing perfectly and then they just stop 00:05:59.240 right and so it seems like there was 00:06:01.360 this large gap between I log into chat 00:06:04.360 gbt or I just use gbd4 from an API or 00:06:07.000 whatever it is and I try something a 00:06:09.039 couple times versus I'm going to sit 00:06:11.160 there and just hammer on this until it 00:06:13.000 works right um and I think if you do 00:06:15.720 that enough you get the intuition for 00:06:17.639 where things are going and how much 00:06:18.919 better they can get um how do we do that 00:06:20.960 from a product standpoint you can kind 00:06:23.000 of think about if you go back to that 00:06:24.800 expand and collapse as like two general 00:06:27.000 themes you need to build productivity 00:06:29.280 tools and what I mean by this is things 00:06:31.800 that are useful for the highest amount 00:06:33.520 of seats right and then you also need to 00:06:36.000 build things that are streamlined 00:06:37.880 vertical workflow from start to finish 00:06:40.039 right and what you can do is you can 00:06:43.160 take those streamlined vertical 00:06:45.319 workflows and you can chain them 00:06:47.360 together to do more increasingly 00:06:49.520 powerful things right and so that's kind 00:06:52.039 of how we think about it internally at 00:06:53.560 the company is you know there's a bunch 00:06:54.960 of bells and whistles and things to add 00:06:56.560 in terms of like sharing collaboration 00:06:58.639 things like that but you also just can 00:07:01.319 basically build certain specific 00:07:03.199 features that will do something from 00:07:04.560 start to finish like I upload all my 00:07:06.879 documents uh and I Target company's 00:07:09.160 documents and it will tell me in 720 00:07:10.720 countries where I need to file antitrust 00:07:13.160 and what do I need to file and what 00:07:14.360 other information do I need and you can 00:07:16.120 take pieces from that and add it to 00:07:17.960 another part of a project how do you 00:07:21.039 organize that effort at Harvey because 00:07:23.879 it's you know like I think you have a 00:07:25.960 very special Instinct and um commitment 00:07:28.879 to going on the the AI bendor I suppose 00:07:31.160 to like figure out what to have going be 00:07:32.840 a long bendor yeah long you know 00:07:34.360 multi-year Bender but you know one one 00:07:35.879 day at a time um uh to devel
