Deal Velocity, Not Billable Hours: How Crosby Uses AI to Redefine Legal Contracting

Contents

Ryan Daniel: I think lawyers are quite good at learning. But in a law firm structure, as much time goes into apprenticeship—it’s a teaching hospital. You don’t actually spend that much time getting really good at teaching because you just do it through reps and reps and reps and reps. And so actually explaining things is something that, like, I think is going to be a very prized skill for not just lawyers, but for any domain experts, but in particular lawyers. And we’re seeing it. Like, when you can make an AI do this thing that you’ve been doing and not enjoying doing, it is like the most magical experience ever. And we’re watching that, like on a weekly basis.

Sonya Huang:  In this episode, we explore a fascinating approach to building services as software with the founders of Crosby, an AI-first law firm focused on legal contract automation. Ryan and John share why they chose to build a law firm rather than legal software and how doing the work themselves creates unique telemetry and feedback loops that traditional evals don’t capture.

Ryan and John also discussed their approach to agent orchestration, from paralegal-level routing to senior associate level contractor review, and share their vision for a Robinhood-like future where anybody has access to high quality legal services thanks to AI. Enjoy the show.

Josephine Chen: John and Ryan, welcome to the show. Thank you so much.

John Sarihan: Thanks for having us.

What is Crosby?

Josephine Chen: All right, first question, maybe. Kick us off. What is Crosby? Tell us a little bit about the company you’ve built.

Ryan Daniel: Great. So Crosby is an AI-first law firm. We focus entirely on contracts. So the theory is that we can automate human negotiations, so getting people to agree on terms. And the best way to do that is with a contract. Contracts, it turns out, are everywhere. They’re your lease, they’re your offer letter, and they’re under most business transactions. And it turns out the best way to do this and to automate it is by building an in-house law firm, understanding how lawyers work and then trying to replace as much of that agentically to get contracts closed faster.

Sonya Huang: And can you talk about the decision to build a law firm versus to build legal software? Because many entrepreneurs in your path—you know, not in the legal space necessarily, but have chosen the path to build software companies. Why build a services company?

Ryan Daniel: So maybe I’ll talk about, like, structurally why we sort of chose this. And then John, you can talk about on a more applied level how we’ve built the team and how that works.

John Sarihan: For sure.

Ryan Daniel: So I think structurally, zooming out a little bit, the market for legal services has always put a huge premium on human capital. And it’s in some ways like a VC firm. I think, like, you’re really investing in the people who work there and getting them better and better and better, and trying to keep those experts or your partners around for as long as possible.

And that made a lot of sense where a lot of that work was very specialized and couldn’t really be offloaded to technology. And there’s a growing sense in the last 30 years, but in particular the last three, that that’s changed. And you’re looking now for a new model that allows you to actually innovate in the kinds of tasks you can offload from humans to technology. And it just has never happened in a law firm before.

But all the raw materials are there, the experts that know how to build things are there. But a law firm partnership, for example, doesn’t allow you to invest in speculative things like building technology because you can’t sell equity, and only law firm partners can take loans that are recourse. So it just doesn’t work. But we had this theory from a structural standpoint that if we get the structure correct—and Atrium proved this five years ago—you would start to see really quick iteration cycles in the way that you can actually innovate. So that’s the theory, at least.

John Sarihan: Yeah. And then putting that into practice, I think having your domain experts sit side by side with your engineers creates such a unique feedback loop, where it’s more than just evals, it’s actually using the product, experiencing it, doing user research and understanding what are the critical workflows, what are the biggest gaps, and actually being able to, you know, recognize and say from start to finish, like, you know, it’s more than just a benchmark, it’s actually like an environment to create that opportunity to, whether it’s innovate on, like, these specific, like, how do you review a contract faster, but also the individual bits and pieces of, like, hey, is this red line exactly correct, or is that the right word choice for what we need here?

Combining software and services

Josephine Chen: I think, you know, historically people have attempted this, so they’ve combined some software as well as the services side together. Tell us a little bit more about why some of these previous attempts might not have worked, but also today operationally, you have AI engineers sitting next to lawyers. How actually does that work? Like, how do you make that successful?

Ryan Daniel: So maybe on that first part of the question: why hasn’t this exactly worked before and what might be different now other than the obvious, which is, you know, the state of the art for AI? I think it actually hasn’t been tried that much, and it’s an extremely daring thing to do. And we’ve seen great innovation in accounting, actually, in management consulting. And when you even look at the corporate entities, fewer and fewer of those entities are partnership structures, and they move to C-corps. And that’s a really good proxy for how much they invest in long-term technology.

And the technology we’ve had in law, we’ve had two or three companies, Atrium and before that Clearspire, actually try this in a meaningful way. And a bunch of things didn’t quite pan out. One was that tight coordination, which I think John you should speak more to, between lawyers and engineers. And we literally in our office have desks staggered: lawyer, engineer, lawyer, engineer, in order to get those feedback cycles and to incent them to really collaborate.

And the other piece I think is—and this is what is truly unique, is that huge chunks of the qualitative, thoughtful work that lawyers are doing each day can actually now be delegated to machines. And this is like, it’s hard to overstate how complex that is and how new it is.

John Sarihan: Yeah. And then putting that into practice, you know, I think that means two things. It means product velocity and it means actually instrumentation. How do you think about metrics? What we’re doing is an operationally heavy business, right? You can actually instrument and think about, we get a contract, it comes in, we send it out really quickly. We need to make sure we’re meeting things like our SLA, which is measured by our turnaround time, how many human touchpoints were involved in actually every contract from start to finish.

And if you can actually understand and say, “Hey, here are all the individual touchpoints, here’s how long each touch point took, and here are the highest value touchpoints we can automate today,” and just start hammering away at those.

I think the other thing when you think about product velocity is it’s more than just, you know, shipping product for the sake of shipping it. It’s do these leading inputs actually lead to customer metrics? Because even though this is an operational heavy business, you need to be thinking about what is the end customer goal that we’re trying to solve for here. Ultimately, our customers come to us because they trust us and we do high quality contract reviews for them.

And, you know, it’s companies like Cursor because they’re growing so quickly. Like, these companies are growing faster than ever before, and we really need to just, like, be able to meet their standards and their bar of, like, what excellence and what speed looks like. And actually taking, you know, all this product velocity, these input metrics, and actually converting those into these lagging indicators of. like, are we meeting their SLAs, are we delivering on time, are we reducing the number of human touchpoints? And then that’s actually how you can drive the impact.

Ryan Daniel: Maybe I’ll just add something that we’ve only started to really get our hands around now. I think a year ago when John and I started working on this idea, we had a strong intuition that structuring this like a law firm was just accessible for companies who wanted to buy some sort of AI-enabled service because there’s so much excitement and people want to try it.

But there’s something subtle, which is legal services, I think economists call it a “credence good,” which means that you only know how good it is after you’ve experienced it and consumed it. And you need an expert to actually tell you the quality. A layperson—or even not a layperson, a sophisticated CEO—often doesn’t know by definition how good their legal work is. And so you want to know that a lawyer is looking at it, and so by structuring it as this law firm, I think we’re taking the most promising aspects of what AI can do, but still putting all of the sort of right safety parameters around it and structures of having an expert tell you this is good quality and this is safe.

Innovating on pricing and packaging

Sonya Huang: How do you think about innovating on the pricing and packaging layer? Like, I think of law as one of the most interesting for the AI wave, because it’s one of the most technology disruptive, but also in terms of pricing model, if you’re billing by the hour, making it more efficient doesn’t really help anything, right? Say a word on how much pricing innovation you’re doing.

Ryan Daniel: Yeah, so this was an early decision that we almost made without thinking too hard about it. And then it’s just been a constraint, which is no billable hour. And that’s like for us, very dramatic. But for us, it was so obvious that it would be an interesting wedge to get people’s attention, and it would align incentives within the company that we didn’t think twice about it.

And so what that looks like today is billing by the document. And I think there’s a few things to say about that. The first is the billable hour, people have been predicting the death of the billable hour for, like, 70 years. It only became popular in the ’50s. It’s kind of novel, but it’s just really durable. And it makes a lot of sense for really sophisticated work where it’s hard to ex-ante predict how much work it will be.

So actually, I think part of what we’re innovating on, which we really could not have guessed a year ago, is being able to predict from time zero how long this piece of work is going to take so that we can price it and still, you know, make sure that we’re doing the work properly, that we align with the value that we’re giving the clients. And, you know, that’s where things get really interesting because you have to predict, you know, how many times will this contract go back and forth—we have to look at it five times, three times, two times—in order to price the amount of—you know, in order to price document properly.

Sonya Huang: Yeah. So interesting.

Ryan Daniel: Yeah.

Sonya Huang: And you’ve been a lawyer before. When you say you guys are automating legal contract work, maybe just take us through what a human lawyer does today when negotiating contracts. And what types of contracts are we talking? Are we talking NDAs? Are we talking merger agreements or employment agreements, everything in between?

Ryan Daniel: Yeah. So many points. I could go on. So cut me off.

Sonya Huang: [laughs]

Ryan Daniel: So right now we focus on NDAs, MSAs, DPAs. I think every tech person will be very familiar with this because you’re selling a B2B product, this is what you’re dealing with. And I’d say NDAs are sort of, from a complexity standpoint, you know, a good deal less complex than MSAs and DPAs. It’s actually quite a step to get to the MSAs and DPAs. You know, an NDA is two pages, an MSA is 15, but they’re probably, what would you say, maybe 80 times more complex? There’s just so many more terms to negotiate. And then you can go all the way up to the merger agreements, which is probably 1000x more complex.

Lawyers today really try to predict, in looking at all the terms in a contract, the right things you should and shouldn’t agree to based on mental models of what’s safe and unsafe, and kind of having to calibrate for your business. We want to get you to sign this contract. And so some of the clients we’re dealing with are signing dozens of contracts a day, right? These are Cursor and Clay and Unify, that are growing so quickly, but you want to insulate the business from too much risk. And I think—so that’s what lawyers do. And there’s a lot of guidelines and general benchmarks and consensus on, like, what is sort of market here. But it’s really, like, ethereal. It exists in the minds of lawyers, and two reasonable lawyers will disagree on what those things are.

And this is a subtle last point that I’ll just touch on, which is, you know, lawyers are essentially these private actors that are building a public good, which is the legal infrastructure, which is all the contracts that ever get negotiated go into this public domain, sort of, of, like, what is a reasonable allocation of risk? And so by automating this—and this is where, you know, I think John gets really interested in the generalizing benchmarks—we can start to actually say quantitatively, “Here’s what that infrastructure should look like.” Not, like, “Here’s my best guess of what’s a reasonable contract,” but, like, we can actually give you a metric or a statistical guess on, like, here’s a risky or safe bet to take.

John Sarihan: Yeah, I think concretely taking the vibes and heuristics inside of a lawyer’s head and converting that into an actual probabilistic quantitative number and saying, like, “Hey, how likely should you be to accept Delaware or California or New York governing law when you’re in the specific negotiation?” I think there’s something innately human about that in terms of that agreement. But at the same time there’s a lot of market data, and you can actually start dissecting and understanding, like, what are the levers you have to allocate risk in your business?

And, you know, circling back to what founders should be thinking about as they’re thinking about these businesses, like, you know, which state you go to court in is like, you know, there’s only 50 states, thankfully for now.

Sonya Huang: [laughs]

John Sarihan: Extending that into, you know, the long tail of what work you could do, I think it’s really easy to confound, like, PMF of, like, your customer is asking you to do this thing and will pay you to do this thing, versus being an actual services business. Like, you need to actually understand what is automatable, what is repeatable, what are the specific signals and levers you have to push that automation forward?

What part AI, what part lawyer?

Josephine Chen: Maybe let’s dig into that, actually. What parts of even the product today, what parts of it are the LLM or AI model? What parts of it are the lawyer? What parts of it are more workflow software? And how does that change over time too?

John Sarihan: Yeah, I think it’s been so interesting in the past, you know, five, ten years, the amount of investment. I’m sure you guys have seen this firsthand, so I’d love your take on this, too. Like, the amount of investment in NLP research, it’s just like, you know, problems that were state of the art five years ago are now like warm-up problems in a Stanford freshman year class.

Josephine Chen: Yeah.

John Sarihan: And I think it’s just been so interesting to say, like, okay, this is what tools we have available to us now. In terms of zooming back out into the specific models and what we can automate today, you know, it comes down to context. When AlexNet and all the first models were starting to do image classification, you had a picture of something and you tried to classify it, all you needed was the picture. You know, now you start thinking about, like, you’re looking at an MRI and you’re trying to classify it. Well, you’re looking at the MRI, but you also might get some additional context from maybe the patient’s history and some other metadata about the patient.

I think language models are now taking that to the extreme. The language models say basically, “Give me as much context as is necessary to understand this problem,” and make some type of decision on it. And to answer your question around the workflows versus the language models, today, it starts with how can you context engineer for a human lawyer, right? How do you give all the right tools and these building blocks to expand what’s available to the lawyer, and actually automate their manual workflows that they were doing? Once you have those building blocks, then it becomes a question, okay, what are the parts that a lawyer is doing that you can now align the language model with and say, “Hey, how accurately can you replicate what behavior this human is doing?” And I think—my hot take here is I think everyone focuses on these general, like, you know, reinforcement, like, RLHF and all these general purpose, like, aligning it to the maximum amount of data available. Really what you want to do in some of these specific scenarios is align it to a specific person, right? Because like Ryan said, even getting two lawyers to agree on one specific contract may require a lot of back and forth and tension, even if they work for the same firm. And if you can actually align it more accurately with one individual, you can start to say, “This is correct. This is wrong,” because that person is internally consistent, even though two people might disagree on that.

Josephine Chen: So you would have almost individualized models for each of the different lawyers. Is that kind of the scenario?

John Sarihan: Exactly. So I think the two theses here are—I think also fundamentally people are not thinking about per customer evals or per customer fine tuning enough. I think per customer fine tuning and evals are a really high ROI way, especially if you have a large ACV, to make your product reach multiple nines of accuracy.

One of the dangerous traps of language models today is they get to 90 percent for basically free. The foundation models have done an incredible job, and we love our partners at OpenAI, Anthropic and Google for giving us such wonderful tools. But one of the dangers is they get to 90 percent, and getting them to 99 or 99.99 is actually extremely difficult. And part of all the levers you have are thinking about, like, what are the ways I can adjust these prompts per customer? What are the ways I can fine tune this model per customer? And then you can actually deliver not just a four-star experience, but a five-star product experience, because really that’s what you’re trying to achieve at the end of the day. But I’d be curious for your guys’s read in terms of, like, the amount of investment for the language models, and how you’ve seen it shape in terms of, like, these per customer, like, these more enterprise focused verticals.

Josephine Chen: I think it partly depends on what, you know, data is actually in distribution for them, because I think whatever data is in distribution, they will—I think they are continuing to improve at a rate where they will probably get there very, very far, versus I think if your point is for a certain enterprise, most of their data, the really interesting data for their contracts you will never have access to as a foundation model company. And we could get access to individuals specifically, and I can create a more RLed fine-tuned model for you that understands your context. Like, I think that makes sense, because it is data that is not in distribution for the large model companies.

Sonya Huang: I agree with that. Good take. You guys have some very discerning customers. I think I heard Clay and Cursor mentioned. Tell us about what they like most about you. Is it that you’re smart AI people? Is it that you can turn around the contract extremely quickly? Is it that you can turn around a contract cheaper than somebody who’s not using AI? What is the customer-facing value prop?

Companies need deal velocity

Ryan Daniel: For these companies—and we’ve really crystallized this over several months—it’s deal velocity. And what we think a lot about is just the acceleration of startups today. And everything is going faster, right? The sales motions are going faster. I mean, you look at Clay, right? Like, they’re, you know, enabling sales teams to move that much faster. The way that you hire, everything. And contract negotiations, which is like the sort of critical piece, it literally is the sort of way you plug in with your customers, the API for business is what we call it, is kind of unchanged for 40 years, basically since the word processor came out.

And so this idea that we can unlock speed in two ways from the time that they send us a contract to the time that they get it back—this is typically the AE or the salesperson—we’re just unlocking their speed that they get back to their client. Or also doing reviews in a more thoughtful way, so that rather than doing five or six back and forths, we can sort of predict, okay, if we just agree to this term and only push back on these three, we’re going to save one turn, so this will be a whole week faster. These are the main things.

And what’s kind of interesting is, as John was saying, sort of quality and taste are these somewhat intangible things that are threshold questions for these kinds of agreements, right? You have to hit some reasonable amount of quality. But I think most great lawyers, the GCs that I really admire and spend a lot of time with, understand their job as business drivers, right? And they are unlocking the growth of their companies, and they position themselves that way. And so that’s the key. And structuring ourselves with lawyers in the loop in a really clever way allows us to have the oversight of lawyers and kind of all the safety and the quality that I mentioned before, but see how much we can really push the limits on right now we do a contract review and median times are at a little under an hour. Can we get that to minutes while still having—making sure that the right terms get in front of the right lawyers at the right times? And these are the fascinating questions.

Josephine Chen: And what does it take to get there, do you think? What needs to happen for that to be possible?

Ryan Daniel: So I think what’s really nice, and I think the reason why John and I were so excited about doing this interview together and why we typically are speaking together, is we had this really unique marriage of, you know, technical expertise, really understanding the cutting edge of AI, as much as I love ChatGPT and, like, what lawyers are really good at. And what’s interesting is, like, when John and I started working together—I don’t know if this is a fair call out. You know, John was constantly asking for evals, and I was like, “That’s fine, we’ll work on this, but I can look at it in 10 seconds and tell you, right? Like, just put it in my hands.”

And lawyers have a very visceral sense, because it’s taste based, based on their training, based on their risk profiles, which is a little unique to each client, just immediately what looks good. And so the more we’ve been able to get prompting tools and actual sort of input and outputs in the hands of the lawyers that we have at Crosby, it’s amazing. It’s just velocity. Like, we actually hired more engineers and lawyers at the beginning, and we were sort of—I felt a little bit like we were sort of revving the engine a little bit without clicking into gear. And then lawyers came, and it was like we stepped into gear. They just had people to give them all the inputs of, “No, no, that’s right. That’s sort of—we should go that direction.” So that’s like a sort of structural thing. In terms of what it takes to get there, I do think that because lawyers provide almost an insurance to make sure that an expert said something looks good, they’ll always be in a loop for a lot of types of legal work. There are, like—and, you know, I can get into hot takes on what kinds of legal work is going to be fully automated, but I think the key here is knowing exactly when to escalate something up to a lawyer and to get their thumbs up, and to know that that liability is covered. And, you know, right now we have malpractice insurance; we take liability for all the work we do. I don’t think we exist as a business if we don’t do that. That’s the key here is we have to be so certain about the quality that we’re able to stand behind it.

Under the hood

Sonya Huang: So cool. Can you give us a peek at what’s happening under the hood on the technology side? You mentioned, you know, you have three beloved foundation model partners. Like, what are you using each of the different models for, and what scaffolding have you built on top?

John Sarihan: Yeah, I think what we found is the state of the art is not specifically tuned for contracts because there’s not enough data in the corpus out there.

Sonya Huang: Yeah, diversity of data is like a bitter lesson. Even if it’s just for the legal domain, you want models that are trained generally.

Ryan Daniel: And these contracts are so hidden away. The best data set is EDGAR. It’s when you attach SEC filings, and that data has just been so overutilized, and doesn’t apply to these smaller companies. So yeah, it’s been an issue.

John Sarihan: I think the other part is in terms of infrastructure, really constantly, like Ryan said, benchmarking this with all the environments we set up for these agents to understand what they’re good at, what they’re bad at, and bringing that back to that quality score that I mentioned earlier. This is why I think every team needs someone, both the domain expert and the engineer. And I think this is why we’ll see more of these vertical AI startups is because having that person in house gives you such a competitive edge, rather than maybe buying a data set from a large company that outsources a lot of, like, labeled data to contract lawyers or things like that.

Ryan Daniel: And then in terms of, like, so, you know, the way it’s set up today, we have lawyers, most of them have big law experience and are pretty proficient in these kinds of agreements, who are driving, right? And so they’re seeing—and this is what’s quite interesting, and to John’s point, the subtleties of the ways you change language within a contract are something that AI seems to struggle with a bit more than we’d imagined. So John and I like to joke the difference between the term “commercially reasonable” and “reasonable” are actually substantively different things to a lawyer, but look very similar in an embedding space. But these are the nuances you really have to pick through.

But what AI is amazing at—so you have lawyers sort of driving and really kind of intervening for the quality on actual edits to a contract. What AI is great at is summarizing, right? So being able to predict the right comment to make, to explain a change. An interesting thing we learned about in May is if you can give a really thoughtful explanation as to why we can’t accept this language, or we really are pushing for that language, the counterparty will sort of understand what you’re getting at and accept it. It reduces turns, right? So it’s worth investing in AI to explain what you’re doing.

Another thing we realized is this is actually very specific to a company. These aren’t legal questions. This is like, you know, some of our clients are not super sophisticated, but you have to understand the fundamentals of the Cursor IDE or the fundamentals of how Clay, you know, uses sales data. It’s just like if you’re a procurement person to know what you’re buying. And this happens in the contract, right? And so you can unlock a lot of speed for lawyers to feed them all this data in the right place, at the right time, in the right parts of the contract. And these are huge unlocks.

As the client experiences it, it just feels like we’ve made it supernatural. You just—by that, I mean very natural. You know, it’s right in Slack. You just tag Crosby. We’ve made a big bet on Slack, I think more and more. It’s just so easy to use. Tag Crosby, send out the document. You get it back within a couple hours with some thoughts, with some comments and with an explanation. And so we’ve built quite a bit on top of that. I think email is the other obvious choice. But we didn’t want any interface. We just wanted to feel like you’re talking to somebody who’s looking at this for you.

John Sarihan: And circling back to the point on infrastructure, you know, getting an agent to be really good at NDAs and MSAs and DPAs and all the—like I said, coming back to the point around the long tail, you really want to one, give agents really good tools, and then two, give them really good context and make sure they’re trained and set for a specific task. And when you’re thinking about the specific task, whether that’s an NDA agent for Cursor and making sure they’re set up for success and actually able to review those contracts really, really quickly. Today we’ve basically built a paralegal agent, which is, like, in charge of routing all the work that comes in from our customers in the same way that a paralegal would at a law firm and then routes it to a human lawyer and makes sure the work is assigned effectively.

The next step is like, how do you think about a junior associate, a senior associate, and a junior partner at a law firm, and think about actually replacing these roles in that way, and the actual specialty jobs that they’re focusing on.

Sonya Huang: Makes sense. Any favorite workhorse models?

John Sarihan: I really like ChatGPT-5. I don’t know.

Sonya Huang: That’s a hot take! Me, too! Like, my teammates make memes for me about it.

Josephine Chen: Wait, why?

Sonya Huang: Yeah, say more.

John Sarihan: I think it’s actually been pretty good in terms of the thinking and pro models for some legal tasks. And Gemini 2.5 Pro has been really good at legal tasks as well.

The promised land

Sonya Huang: Super interesting. Do you think that the path to the promised land for you will be kind of RL tuning to each individual law firm and lawyer, or do you think it’s in prompting? Like, any religion or any gut instinct on what’s going to take you to the promised land for automating these contracts fully?

John Sarihan: Well, I think, like Ryan said, the thesis is how do you still keep that human touch for, like, psychological safety? I think the promised land comes from really good context engineering and providing it to the—right? Because what great product counsels will do and what great commercial counsels will do is keep just a ton of information in their working memory about here’s how this company works, here’s the background on this specific playbook, here’s how this NDA last time went when we negotiated against this company. Loading all that into the working memory is going to be really important. I think I’m really optimistic about a lot of these reinforcement learning techniques. So reinforcement fine tuning was something we experimented with early on for comment generation, like Ryan mentioned, and have experimented with it for a lot of other opportunities as well. And so pretty optimistic there.

Ryan Daniel: My take is, like, let the lawyers cook, you know? Get them writing prompts.

John Sarihan: [laughs]

Sonya Huang: Do your lawyers cook? Say more.

Ryan Daniel: Yeah.

Sonya Huang: Are they writing prompts?

Ryan Daniel: We speak to so many lawyers, you know, who are like, “All I hear about is AI. I want to learn more. I’m curious, I want to be part of it.”

Sonya Huang: Yeah.

Ryan Daniel: And you know, you’re like, “Have you used it before?” “No.” And we’re like, okay, how’s this gonna play? And they’re like—it’s like the look on their face. I’m thinking of one lawyer particular on our team. The look on her face the first time she saw some of the tools we were doing, and she just—like the lights turned on. Like it was so neat. And so, like, to get them access to actually write prompts, give instructions, teach all these things that are, like, in your head, but you wouldn’t even know how to explain them to someone. And it’s actually quite complicated to try to tease these things out and explain them.

I think lawyers are quite good at learning, but in a law firm structure, as much time goes into apprenticeship—it’s a teaching hospital. You don’t actually spend that much time getting really good at teaching because you just do it through reps and reps and reps and reps. And so actually, like, explaining things is something that I think is going to be a very prized skill for not just lawyers, but for any domain experts, but in particular lawyers. And we’re seeing it. When you can make an AI do this thing that you’ve been doing and not enjoying doing, it is the most magical experience ever. And we’re watching that, like, on a weekly basis.

Josephine Chen: What’s the coolest thing a lawyer has created? And maybe just, like, describe how you even create a culture where lawyers are prompting themselves, you know, they’re actually playing against themselves.

Ryan Daniel: So I’ll give a big shout out, like, we have to give kudos to Harvey for building, like, such a center of gravity to show that, like—you know, they invented this job title, I think, called Applied Legal Research, which is brilliant. And, like, that is what lawyers are doing.

And you create this culture that really praises that. You know, it doesn’t just praise—you know, you look at lawyers at law firms, they’re measured really on one vector, which is hours billed each year. And so we give a lot of public praise for being sort of meta aware of the work you’re doing and trying to measure it out. So we have one lawyer who, like, you know, kind of mentioned, he thought we could be doing things better. And I said, “Tell me a bit more.” And a week later he built this, like—I taught him Miro, which was like a huge a-ha moment for him. And he made this huge process map that, like, we had to get it printed professionally, it’s so big. And, like, it was one of the best moments for all the engineers.

But, like, we were so excited about it. And so we created this culture, like, we really praise—you know, we’re doing a lot of repetitive, the same kinds of work, but there are ways that we can just, like, step out of that and think bigger. And so you really incentivize lawyers, again because your incentive is to get the TTAT, the total review time, lower and lower to be the ones driving that. And so it’s been neat. And, you know, there’s so many other things that we’re just beginning to scratch the surface of, like organizational design, structuring your team so that there’s both lawyers and engineers all towards the same outcomes, and so they all have to collaborate together. But these are the keys and these are the really hard things.

Building in New York

Sonya Huang: Can you talk about New York? You’re building an AI company in New York. That is contrarian. Tell us more.

Ryan Daniel: It’s not contrarian to us. [laughs]

John Sarihan: I think the New York tech scene has gone through a lot of different arcs. The first arc of the early 2000s, a lot of these ad tech companies that were emerging after the dot com boom like AppNexus and all these other companies that just exploded. And then there was a second wave of all these deeply technical engineers that came out of these ad tech companies spinning off and creating these more dev tools like MongoDBs of the world. In parallel, there were a lot of also trading firms in New York, right? So you can think of the Jane Streets, the HRTs, the Citadels.

So the New York tech team had this as, like, its base framework. What happened in the last five to ten years is, you know, more and more senior engineers were coming out of those places to the point where a lot of what building a startup is about is actually knowing what a great growth trajectory looks like and understanding how do you not just go from, like, one to ten but zero to one. And so these were great breeding grounds. And now we’ve seen a ton of companies start spawning out of, like, these, whether it’s in dev tools and DevEx or thinking more—you know, I think a really great example is Ramp where I remember one of the first—Karim told me this story about how they had a YC company, it was called Paribus and it was based in New York. They said that I think they might have been the only YC company in New York in 2013 or something like that.

The reason it’s so interesting is because they had a candidate who applied, extremely talented developer, incredible resume. And they asked him why did you apply to work here? And he said, “One simple reason: I filtered by New York City on YC and you were the only company there.” [laughs]

And so we’ve come a long way since then, I think. You know, Ramp has been just an incredible place, and there’s many other companies in New York now coming out of that as well, I think, where you can take a lot of great young talent, teach them from these people who worked at these places like AppNexus and all these other places of, like, other startups that have scaled, apply those learnings and create a truly generational company like Ramp is.

There was an interesting survey—I think Pat talked about this as well, or you might have mentioned it, where there was a survey from Neo or someone else where most new grads wanted to be in New York in 2021, 2022. And it’s so interesting where if you just follow that as a leading indicator of, like, where do all the really smart, hungry, high slope, young engineers want to be? And I think that was New York.

Sonya Huang: Yeah. It also allows you to break out of the AI echo chamber a little bit and do things a little bit differently. And just hearing how you guys speak, it’s refreshingly first principles.

Ryan Daniel: I do think, like, you know, we’re probably here every few months for various reasons, and I feel like we come here to dream, and then we go back to New York to build. And the subject matter expertise there is so dense in finance, in the creative fields.

John Sarihan: In healthcare.

Ryan Daniel: Healthcare. Obviously in law. And you sort of distill, like, all the really crazy ideas—like, you come here and you’re like, it’s not AGI, it’s ASI. Like you learn all this, right? And then you go back and then you really apply it, and you sort of like—you know, like the water goes through the sand and it comes out with something that kind of just works. And that’s been really special to be part of.

John Sarihan: Yeah. And I think for us, really, you know, taking these domain experts with really deep expertise and combining them with these, like, high slope new grads as well as these people who are, like, either experienced founders or want to be founders again—our entire founding engineering team, like, is all either previous founders or wants to be founders—I think that mentality is really what it takes, because in some ways, like Ryan said, you’re building these pods and you’re working with one or two lawyers and your customer is sitting right next to you. It’s such a unique opportunity in terms of the amount of dev cycles and product loops you get. And I think that’s really special.

Josephine Chen: I mean, your name itself kind of echoes your tie to New York. Maybe tell us a little bit about the story of the name.

Ryan Daniel: Well, there’s a lot of myth and a lot of lore over the name of Crosby. One of the myths is that John and I were going on these long searching existential walks back last summer through SoHo. And we kept finding ourselves on Crosby, and it’s a beautiful old street in a very modern neighborhood. And it kind of to us spoke about being an essentially New York company that has all the best artisanship of the old and all the modern steel glass of the new. And it combines it nicely. But that’s just one. It could be after the hockey player, also.

Josephine Chen: Could be after the hotel. Could be anything. [laughs]

John Sarihan: Our office manager’s dog was named Crosby. Yeah.

Ryan Daniel: Really?

John Sarihan: Yeah, her dog.

Josephine Chen: Maybe tell us a little bit more also in terms of just having built in New York, like, what is different about the company culture, actually? Like, how do you think it is distinct from a typical AI company maybe here in SF?

Ryan Daniel: I do think—I don’t know why this would be a New York thing, but I’m seeing it more and more in some of our friends’ companies, other Sequoia companies in New York. There’s a huge bias to starting things, to being founders. And the way that we’ve been able to convince people to join us is by saying, “This will be your stepping stone for four or five years.” And we have to live up to that, by the way. Like, it’s not trivial. We have to make sure that, like, when we get invited to a founder dinner, we send someone else in our stead, which I think annoys people, but it’s pretty awesome for them. But what’s interesting is in this type of company, on the application layer, as John mentioned, you need that level of autonomy and agency and creativity of that founder mentality of, like, “This is my thing, and I’m going to figure how to build it with my pod of lawyers, and I’m going to solve this and then I’m going to come back.” And so I don’t know why, but we keep seeing more of this. I do think Ramp’s a big part of the story. I think Ramp has become a founder factory, and sort of culturally that permeates all throughout New York, and we feel it. So I’m actually really excited and I’m curious to watch the next four or five years of, like, all of these kind of the next wave of folks coming out and starting things.

John Sarihan: Yeah. I think one other uniquely New York thing is just the emphasis, it feels like, on design. You know, Ramp is a product that’s really prided for its, like, great design and product taste. I think a lot of that can be traced to just the depth and creativity levels at these design agencies and other great firms. One last thing: people in New York like pizza. I feel like it’s really hard to get a proper New York slice, or any slice in San Francisco.

Josephine Chen: That’s how we celebrate it.

John Sarihan: That’s how we celebrate.

Josephine Chen: Pizza.

Ryan Daniel: So we have real two-pizza teams in New York. Not like …

What gets completely automated?

Sonya Huang: Okay, we need the hot takes. What type of legal work is going to be completely automated?

Ryan Daniel: So legal work, there’s about 11,000 law firms, which account for eight percent of all law firms in the US, that account for about 75 percent of revenue. That’s still a lot of law firms, but there’s this huge tail of these other, you know, 92 percent that make up the rest. And those 92 percent focus on individuals, helping you with child support payments and leases and, like, all the sort of human things. And they’re, like, horribly underserved by lawyers. It’s almost immoral.

And we talk about all the time in law school, and everybody has really lofty dreams to kind of do something about this and it’s not really being touched. And I think this work will be automated entirely because the alternative is nothing. And I suspect that anybody right now can negotiate a landlord kind of lease with ChatGPT today. And I think a lot of people do. I think that the corporate law firms are here to stay for quite a while, and those jobs are quite safe. So it’s like sort of a net new type of legal skew that’s going to be totally automated. But it’s not that we’re taking a lawyer’s job and automating it, it’s just nobody’s doing it today.

Sonya Huang: That’s net new market creation.

Ryan Daniel: Totally.

Sonya Huang: Why isn’t it just ChatGPT that does it, though? I already kind of turn into it for legal advice. Maybe bad strategy, but I …

Ryan Daniel: I think for a lot of this work it is. I think again, the alternative is nothing; the alternative is having nobody look at it. And my hope is that if you can add so much leverage to lawyers that, like, rather than looking at two contracts an hour, they can look at 500, like, shouldn’t we be able to have more and more people getting serviced? Like, that, I think, when we talk about the—our mission of Crosby is very lofty. It’s building a better legal infrastructure with technology. This is what we’re talking about. This is just so leveraging the lawyers that we have. It’s an optimistic hot take. I think it’s a golden age for what lawyers are going to be able to do in the next five, ten years being unlocked by this.

North star metrics

Sonya Huang: I love that. Do you have any north star metrics that you run the company on?

John Sarihan: Yeah, we run on total turnaround time, which means time in to time out, how long does the contract take across all of the back and forths? So like Ryan said, you know, a contract negotiation might have five back and forths or two back and forths. If you add up all the time that Crosby has spent looking at a contract, that’s our north star metric. And that’s very counterintuitive, right? Most law firms are not trying to maximize that number, but that’s how they earn their billable hours. For us, it’s about aligning incentives. We want to not only do a faster job on each individual turn, it’s also about how do we do a better job, how do we understand the right negotiation levers for the business? I’m sure you all have negotiated your fair share of term sheets that are really high velocity coming back and forth. For us, it’s really thinking about, like, how do we minimize those back and forths and also get faster every time?

Sonya Huang: How do you make that incentive aligned for your customer? Like, if you go back to the term sheet example, the fastest way I have of getting a term sheet signed is to increase the post money valuation. So how do you make sure that you’re not giving in on your customers negotiating leverage in order to kind of maximize that north star?

John Sarihan: Exactly. It comes back to guardrail metrics as well. And so we really pride ourselves—we have, I think, one thing that most, you know, startups that think about these vertical AI opportunities is they should have some type of team focused on either environments for their AI agents or quality metrics for their agents. And the reason that matters so much is because it’s very easy to get misaligned incentives of, like, taking shortcuts, or doing work that’s, like, halfway there. We have people on the team, both lawyers and engineers, only focused on how do we make sure quality is consistently meeting our customers’ needs.

And that means not just doing good legal work in an objective sense, but also, hey, is this aligned with our customer’s risk profile? Is this, you know, in the best interest of the customer’s negotiation? And distilling all that down into a guardrail metric, combining that with the total turnaround time, we call it TTAT. And then I’ve learned that also from Ramp, you know, if you have pithy metrics that you can really, like, rally behind that are fun to say, it actually does help just because people start—yeah, we just came up with it, you know, at one meeting, and now the entire office goes around saying “TTAT.” We also called this other metric, it’s called HuRT, which stands for Human—hu—review time—RT. And so we’re also trying to reduce the amount of HuRT.

Josephine Chen: I like the puniness of this.

John Sarihan: And I think it’s funny because I think metrics, you know, you can talk about them from a product perspective, but there’s something very innately human, and you want it to feel empathetic, right? Like, reducing HuRT feels more value aligned with how our lawyers should feel. And also the same way, like, quality is so aligned with what our customers are expecting from us.

Ryan Daniel: I think one thing maybe I’ll add: the other ways that you have to increase leverage in a negotiation, you know, in a fundraise is, like, set deadlines, like, increase the pressure. Like, there’s a lot of ways. And I think contracts are so interesting because they’re so human. Fundamentally, it’s just like an abstraction of a human-to-human conversation about, like, what can we agree on?

And so last summer when we were kind of researching the idea, I’d actually spent time in India—John’s heard this story countless times. And there’s a lot of offshore legal services that do sort of back office, very routine contract negotiations for Fortune 500 global companies. And they just take five rules and apply those to a contract, and just kind of go back and forth. And, like, negotiating against them is like negotiating against the wall. They just—you know, they kind of don’t get creative. But I had this one guy, and I said, “So, like, what do you do if you get stuck?” He’s like, “I have this really creative thing I do. I just call them and tell them that they’re not going to want to say no to, you know, this company, and we have a really long life and we’re going to want to interface.” And I was like, “Are you allowed to do that?” He said, “I don’t know. I just started doing it and it works.” Like, it was so human, and it made me so appreciative of the human element of these back and forths. And so I do think, like, we’re unbelievably optimistic about how much time we can reduce, and how much more efficient we can make the transactions between entities. But also, there’s an essential role for those counterparties to have in interfacing with each other to have what we call meeting of the minds, right? To really get to agreement.

Sonya Huang: Totally.

John Sarihan: And at the same time, I also wonder if we had an AI agent with some type of reinforcement learning, if it just would reward hack and realize that picking up the phone would a hundred percent convert if I give it a tool call for phone calls. But who knows?

Agent vs agent

Sonya Huang: Well, how do you think this plays out though? Like, let’s say—I imagine your agent has never negotiated against another one of your agents before in the field. But, like, eventually I’m guessing people will have their agents negotiating NDAs on both sides, right?

Ryan Daniel: I think this is the most beautiful essence of what we hope we can achieve here, is being able to capture each party’s preferences with their own sets of agents, and have them simulate the negotiation and be able to show an auditable record of here was the first back and forth and then the fifth and the seventh and here’s where we netted. And what do you think? Is that reasonable? And that’s your starting point. And these are fundamentally collaborative negotiations; these are not adversarial. And so the goal here is actually to be the collaboration platform to get people to agreement faster.

Sonya Huang: Hmm, so interesting. So each agent will kind of have a sense of everyone’s risk preference, their bottom line, their tolerance for waiting out the negotiation.

Ryan Daniel: And I want the agent that’s going to, like, bang its head on the table.

Sonya Huang: [laughs] Yeah. And, like, that’s the one that you’re going to spend more compute on.

Ryan Daniel: The high inference agent.

Sonya Huang: Yeah.

John Sarihan: And yeah, I think fundamentally, you know, partitioning the data, making sure, like, this agent doesn’t have access to the other agent’s data, and making sure, like it’s just really taking what law firms have already figured out of, like, you set up a wall between different teams and say, “Negotiate against each other.” You can apply that even in an easier way in a system setting, because you just don’t need the agents to talk to each other unless through their specific guardrails.

The legal industry in 10 years

Sonya Huang: Great. I’d love to bring it home by asking a bit about the future of how you guys see Crosby, and how you see the legal industry playing out. What do you think the legal industry looks like in 10 years?

John Sarihan: You know, I think coming back to the point on taste, increasing the leverage of these senior partners at law firms, I get a little worried about the junior associates and the paralegals and the opportunity for them. I can imagine it just looks like a senior associate managing an army of agents who can increase their leverage. Maybe for a personal injury firm, it’s managing intake and someone managing actually drafting the demand letters and then actually reviewing them. The main job of the person is to show up in court, and everything else is handled by an agentic system. But I’m curious for your take.

Ryan Daniel: All right, I’m going to try to do this quickly. In the 10-year period from 2007 to 2017, in-house legal teams grew 200 percent and law firms grew 30 percent. And then again from 2014 until 2024, they went from about 320,000 to 440,000. In-house teams are really, really expanding, and the narrative seems to be that legal work is getting more complex, and companies need much bigger in-house teams. And in-house teams have a lot fewer constraints than law firms. They can have not just lawyers, they can have paralegals, they can have legal operations, they can have different types of work—specializing. And they’ve become really creative. And it’s this kind of huge change that’s been going on quite quietly.

My hope is that—you know, because these are not companies that specialize in legal services, this is the legal function, you’ll see more and more specialized companies that are AI first, because it’s just so obvious, that are truly changing the way legal services get delivered. We are, like, a small part in this, and I’ll go back to what I said. This is the golden age. I think this is, like, we’ve never seen true innovation in the legal market. And we’re just at the beginning. I think things are going to look very different.

Josephine Chen: And maybe given that future, what should somebody in law school today do? Should they drop out? What should they be doing?

Ryan Daniel: I think this will go back to the fabric of our company. I think you want to question absolutely everything. I mean, like, question the way your professors tell you to write footnotes. Like, is that really necessary? Question all—there’s so much dogma. And then balance that—like, it’s so fragile—with understanding the brilliance of legal academia. This is what makes society run. My professors at Stanford truly are some of the best mentors I’ve ever found and some of the smartest people. And so don’t be arrogant, but question everything. And come take time and do an apprenticeship with Crosby. We’ll teach you how to prompt.

Josephine Chen: [laughs]

Ryan Daniel: But, like, it’s important. Like, these are the skills that I think are really going to matter, is how to harness all the power of AI. And there’s so much room to change things, and there’s so many things that are accepted as just facts that can all be permeable. And it’s hard to see when you’re in it.

John Sarihan: Yeah. I think—I won’t say which company, but I went to a conference. It was all general counsels—imagine Fortune 500. I won’t name the conference because I would like to go back, but I sat next to this gentleman. He was the general counsel of a very large telecoms company—maybe one of our phones runs on this telecoms company. And I said, “Hey, have you, like, ever tried any of this ChatGPT or AI stuff?” And he said, “No, of course not. My CEO keeps telling me to. I refuse. I do not want to try it.” And he’s just—you know, I think to Ryan’s point, like, there’s arrogance on both sides of, like, you know, questioning everything, but also, like, you know, being stuck in the old ways as well is dangerous.

Josephine Chen: Makes sense. Well, I think question everything and try everything seems to be a good theme for all of us. So thank you guys again for coming on. Appreciate it.

John Sarihan: Thanks for having us. It was really fun.

Mentioned in this episode:

Mentioned in this episode:

• Atrium: Failed legal startup founded by Twitch co-founder Justin Kan
• Clearspire: A pioneering but also failed legal tech firm founded in 2008 that created the Coral software platform for law firms 
• Data processing agreement (DPA): GDPR-mandated contract between controllers and processors. Crosby handles DPAs as part of B2B contracting, along with NDAs and MSAs
  
• Credence good: Economic term for services whose quality is hard to judge even after consumption. Used to explain why legal buyers value lawyers-in-the-loop and malpractice coverage.