Crosby Bills by the Document, Not the Hour, and Turns Contracts Around in Under an Hour

Crosby is an AI-first law firm that does one job: get people to agree on terms, which in practice means contracts. The current diet is NDAs, MSAs, and DPAs, the stack every B2B company lives in. An NDA is two pages; an MSA is fifteen, and Ryan Daniels puts the complexity gap at something like 80 times, with merger agreements another thousand times beyond that. Clients include Cursor, Clay, and Unify, companies signing dozens of contracts a day, growing faster than the legal function that is supposed to keep up.

The structural reason they became a firm rather than a software vendor is that legal services have always priced human capital the way a VC firm does: invest in experts, keep the partners. A partnership cannot sell equity and only partners can take recourse loans, so it cannot fund speculative technology. Ryan treats the shift from partnerships to C-corps in accounting and consulting as a proxy for long-term tech investment. In law, the serious attempts were Atrium, Justin Kan’s firm that shut down around 2020, and Clearspire before that. Get the structure right, he argues, and iteration speed changes.

John Sarihan’s version is operational. Desks in the office are staggered lawyer, engineer, lawyer, engineer, because the feedback loop is not a benchmark. It is someone using the product, feeling a bad redline, and saying the word choice is wrong. The work is instrumented as an operations business: a contract comes in, it has to go out against an SLA, and every human touchpoint is timed so the team can hammer the highest-value steps that are automatable today. Product velocity only counts if those input metrics turn into lagging ones customers feel: on-time delivery, fewer touches, reviews that still meet the bar.

The subtle reason the law-firm wrapper matters, Ryan says, is that legal services are a credence good. You often know the quality only after you have consumed it, and you need an expert to tell you it was good. A sophisticated CEO, by definition, cannot evaluate the work. Buyers want to know a lawyer looked. Crosby puts the promising parts of AI inside that existing safety structure, then buys malpractice insurance and takes liability on every file. “I don’t think we exist as a business if we don’t do that,” Ryan says later. Without someone to stand behind the output, the company should not exist.

The pricing decision was almost made without thinking, then became a hard constraint: no billable hour, ever. They charge by the document. People have been predicting the death of hourly billing for 70 years; it only became popular in the 1950s, and it still makes sense for work whose scope you cannot predict in advance. Crosby’s unexpected R&D problem is the opposite. From time zero they have to forecast how long a matter will take, how many turns it will take, two or three or five, so the price can cover the work and still match the value to the client. That prediction problem is where the product gets interesting.

What lawyers actually do on these files, Ryan says, is hold ethereal mental models of what is safe, unsafe, and “market,” then calibrate to the business that needs the deal signed. Two reasonable lawyers will disagree. Lawyers are private actors building a public good: every negotiated contract feeds a shared sense of how risk should be allocated. Automating that, John adds, is how you turn vibes into numbers. Should you accept Delaware, California, or New York governing law in this negotiation? The answer can become a probability instead of a feeling, which is the beginning of a quantitative picture of legal infrastructure rather than a partner’s best guess.

John’s map of the stack starts with context engineering for a human lawyer: tools and building blocks that expand what the lawyer can see, then a question of how faithfully a language model can copy a specific person’s behavior. His hot take is that the industry over-indexes on RLHF-style alignment to the largest possible dataset. Two lawyers at the same firm can fight over a clause. One lawyer is internally consistent. Align to that person and you can finally label a redline correct or wrong. The host asks whether that means individualized models per lawyer. John’s stronger claim is adjacent: people are not thinking hard enough about per-customer eval and per-customer fine-tuning.

The trap of current foundation models, he says, is that they get you to 90% for basically free. OpenAI, Anthropic, and Google have done that job. Getting to 99 or 99.99 is extremely hard, and the levers are per-customer prompts and fine-tunes, especially when ACV is large. The host agrees on the data point: whatever is in distribution, the labs will keep improving; the interesting contract data of a given enterprise will never be. That is Crosby’s edge if they can RL-fine-tune on context the labs cannot see. State of the art is not tuned for contracts anyway. The best public set is EDGAR, already overused, and it does not look like the everyday paper of a small company.

What those customers actually pay for crystallized over months: deal velocity. Startups’ sales and hiring motions have sped up; contract negotiation, “the API for business,” has been roughly unchanged for 40 years, since the word processor. Crosby unlocks speed in two ways. Raw turnaround, so an AE gets a marked-up file back and can reply to the counterparty. And smarter negotiation: agree to this term, push only on these three, save a whole turn, save a week. Quality and taste are threshold questions. The GCs Ryan admires position themselves as business drivers. Median time is already a little under an hour. The live question is whether they can get to minutes while still putting the right terms in front of the right lawyers at the right time.

The marriage that makes the target plausible is John’s demand for evals and Ryan’s 10-second visceral “that looks wrong.” Early on they hired engineers and lawyers and felt like they were revving without being in gear; the lawyers arriving was the clutch. Lawyers will stay in the loop for a lot of this work because they are, among other things, insurance that an expert said it looked good. The craft is knowing when to escalate, get a thumbs-up, and know the liability is covered. Under the hood, agents are benchmarked in the environments the team builds, which is why John wants a domain expert and an engineer on every vertical AI team rather than labeled data bought from an outsourced shop.

Language is where models still stumble. “Commercially reasonable” and “reasonable” are substantively different to a lawyer and almost the same in embedding space. Lawyers drive the actual edits. Where AI shines is comments: predicting the right explanation for a change. A finding from about May: a thoughtful explanation of why language cannot be accepted often gets the counterparty to yield, which reduces turns. They invested in that on purpose, including early experiments with reinforcement fine-tuning for comment generation. The explanations are not purely legal. They have to know how the Cursor IDE works, or how Clay uses sales data, so a procurement person understands what they are buying. Delivery is deliberately interface-free. Tag Crosby in Slack, send the document, get it back within a couple of hours with comments and an explanation. It is supposed to feel like a person looking at it for you.

Today they have a paralegal agent that routes inbound work to a human lawyer the way a paralegal would. Next is the associate-to-junior-partner stack, replacing those roles’ specialty jobs one at a time. John’s workhorse models are GPT-5, which he likes for thinking, and Gemini 2.5 Pro, which he finds strong on legal tasks. The host calls GPT-5 a hot take and admits teammates make memes about her agreeing. The promised land, John says, is still context engineering: loading a great commercial counsel’s working memory, how this company works, how this playbook was born, how the last NDA with this counterparty went. Ryan’s cultural instruction is simpler. Let the lawyers cook. Get them writing prompts.

They meet plenty of lawyers who say they want to be part of AI and have never used it. One Crosby lawyer’s face, Ryan says, when she first saw the internal tools, was the lights turning on. Teasing out what is in someone’s head and they would not know how to explain is hard. Law firms are teaching hospitals. Time goes into apprenticeship by reps, not into getting good at teaching. Explaining will be a prized skill for domain experts, lawyers especially. When you make an AI do the thing you have been doing and not enjoying, it is “the most magical experience ever,” and they watch it weekly. Harvey gets a shout-out for inventing “applied legal research” and building a center of gravity around it. Crosby praises lawyers who step outside billable-hour logic: one of them learned Miro, mapped the process so large it had to be professionally printed, and the engineers were thrilled. The incentive is explicit. Drive TAT down.

Building an AI company in New York is, to the host, contrarian. To John it is just the next arc of a scene that already had ad tech after the dot-com boom, then MongoDB-style developer tools, then Jane Street, HRT, and Citadel as a deep engineering base. The last five to ten years produced senior people who know what a growth trajectory looks like. Ramp is the example. Karim’s earlier YC company, Paribus, may have been the only New York YC company around 2013. A candidate applied for one reason: he filtered YC by New York City and they were the only hit. Ramp is now a founder factory, and that culture, Ryan says, permeates the city.

The host’s line, which Ryan adopts, is that they come to San Francisco to dream and go back to New York to build. Subject-matter density in finance, media, healthcare, and law is the filter. Crazy ideas get poured through that sand and come out as something that works. Crosby’s founding engineers are previous founders or people who want to be. The pitch to join is that this is a four- or five-year stepping stone, which they have to live up to, including sending someone else to founder dinners in their place. Application-layer companies need that pod-level autonomy: this is my thing, these are my lawyers, the customer is at the next desk. Even the name is New York lore. One myth is long walks through SoHo that kept ending on Crosby Street, old artisanship against new steel and glass. It could also be the hockey player, the hotel, or the office manager’s dog, which is in fact named Crosby.

Ryan’s hot take on full automation starts with market structure. About 11,000 firms, 8% of U.S. law firms, take roughly 70 to 75% of revenue. The other 92% serve individuals: child support, leases, the human pile, “horribly underserved,” “almost immoral.” That work, he thinks, will be automated entirely because the alternative is nothing. Plenty of people already negotiate a landlord lease with ChatGPT. Corporate Big Law jobs, in his view, are safe for quite a while. This is net-new market creation, not stealing a partner’s book. The host asks why it isn’t just ChatGPT. For a lot of that tail, Ryan says, it is, because nobody else is looking. Crosby’s loftier mission is leverage: if a lawyer can review 500 contracts an hour instead of two, more people get served. He calls it a golden age for what lawyers will be able to do.

The operating system of that leverage is metrics with names you can shout across the office, a habit they borrowed from Ramp. North star is total turnaround time, in to out, summing Crosby’s time across every back-and-forth. Most firms are not trying to minimize that number; it is how they earn the hour. The host’s objection is the obvious one: the fastest term sheet is the one where you give on post-money. Guardrails exist so TTAT does not become giving away the deal. A dedicated mix of lawyers and engineers watches quality against the customer’s risk profile and negotiation interest. They also track HuRT, human review time, and they like that “reducing hurt” sounds like something a lawyer should want. Fun-to-say metrics, John says, actually change whether the office rallies.

Contracts are still an abstraction of a human conversation. Last summer in India, researching the idea, Ryan watched offshore teams apply five rules to Fortune 500 paper and negotiate like a wall. One person had a different move: when stuck, he called and said the counterparty would not want to say no, they had a long life together. Was he allowed? “I don’t know. I just started doing it. It works.” That story is why Ryan still wants a meeting of the minds, not only a faster markup. John’s joke is that an RL agent with a phone tool would reward-hack calling, because calling converts. The endgame they actually want is each party’s agents capturing preferences, simulating the negotiation, and producing an auditable record from the first turn to the seventh. These are collaborative fights. The product is a collaboration platform to agreement, with walls between agents the way firms already wall teams.

Ten years out, John worries first about juniors and paralegals. He can picture a senior associate managing an army of agents: intake, demand letters, review, with the human’s main job showing up in court. Ryan’s numbers are about in-house, not the firms. From 2007 to 2017, in-house legal teams grew 200% while law firms grew 30%. From 2014 to 2024, in-house headcount went from about 320,000 to 440,000. Those teams have fewer constraints. They can hire paralegals, legal ops, specialists. His hope is a wave of AI-first specialty companies that change delivery, of which Crosby is a small part, and that the legal market is only at the beginning of real innovation.

Asked what someone in law school should do, and whether they should drop out, Ryan goes back to the company fabric. Question everything, including how professors tell you to write footnotes. There is enormous dogma. Balance it with the brilliance of legal academia, which is “what makes society run.” Do not be arrogant. Then come apprentice at Crosby; they will teach you to prompt. John will not name the conference, because he wants to go back, but he sat next to the GC of a large telecom who refused to try ChatGPT even though the CEO kept asking. Arrogance, he says, exists on both sides: questioning nothing, and being stuck in the old ways. The host’s closer is the usable version of the hour. Question everything and try everything.