An all-in-one AI legal assistant built in Palo Alto. Research, drafting, document review, and contract analysis - in one workflow, with citations attached.
Somewhere in Palo Alto, a lawyer opens a browser tab, drops a thousand-page deposition into a chat window, and asks a question in plain English. The answer comes back in seconds - with citations, with passage numbers, with the exact paragraph it came from.
That is Paxton AI in 2026. Not a research database. Not a fancy autocomplete. An assistant that has read the file and is willing to argue about it. The mission, stripped of pageantry, is unromantic and useful: take the parts of legal work that look like reading and writing very carefully, and let software do them faster without doing them worse.
This is the story of the company that decided, in 2023, that large language models had finally crossed the line where they could be trusted with law - if you built around them carefully enough.
Tanguy Chau holds a master's, a PhD, and an MBA from MIT - a vita that, in most other careers, would suggest a person comfortable in laboratories rather than legal hallways. Before Paxton, Chau was a venture investor at Formation 8, where in 2015 he put money into Ironclad, the contract-management company. He noticed something the rest of the legal industry was still arguing about: AI was going to read documents better than most associates, and soon.
He spent a decade thinking about that thesis from the buy-side. Then, in 2023, he stopped thinking and started building. He brought in Michael Ulin, a technologist whose background spans AI, data science, and market intelligence, as co-founder and CTO. Ulin handles the engine room - the retrieval, the grounding, the benchmarks. Chau handles the rest.
Paxton is not built around a model. It is built around a workflow. Research, drafting, review, analysis - the four verbs of a billable hour - each get a dedicated lane inside one chat surface.
Real-time search across federal and state statutes, regulations, and court rulings. Every answer cites its source.
Contracts, clauses, demand letters, internal memos. Generated with authority links built in.
Ask a question of a 400-page filing. Get the answer, the passage, and the page number.
Surfaces risk language, missing clauses, and inconsistencies across stacks of agreements.
Consolidates line items and totals for demand letters and settlement negotiations.
A firm-specific repository so Paxton answers using your past work, not just public law.
For most of the 2020s, legal AI looked like an enterprise sales pitch. Pricey contracts. Long pilots. Two-year procurement cycles. The product roadmaps were tuned to whatever the AmLaw 20 wanted to see in a quarterly business review.
Paxton's customer base tells a different story. Yes, several of the country's twenty largest firms use it. But so do solo practitioners running plaintiff-side personal-injury work out of a single office, and mid-sized regional firms whose IT budget would not survive a Thomson Reuters renewal. The platform is priced and packaged for both ends of the market - which is why the funnel grew 14× MRR in nine months.
That is not a marketing trick. It is a deliberate product choice. If your AI is only useful to firms with a procurement team, you are leaving the actual majority of the legal profession on the table.
Paxton publishes its results on a hallucination benchmark developed by researchers at Stanford. The reported number is 94% non-hallucination. That is not a number any vendor publishes lightly - it invites comparison. Paxton seems to enjoy the comparison.
Sources: company filings, Series A press, Stanford benchmark study (Jul 2024).
Tanguy Chau and Michael Ulin start Paxton AI with the bet that generative models have crossed the line where they can be trusted with legal work - if grounded carefully.
Early backers WVV Capital, Kyber Knight, and 25madison commit to the platform. Paxton ships its first wave of drafting and research features.
Paxton publishes a study using Stanford-derived benchmarks and reports a 94% non-hallucination rate. The number becomes a reference point in the legal AI space.
Unusual Ventures leads the round, with returning investors Kyber Knight, 25madison, and WVV Capital. Total funding to date: $28M. The customer base crosses 1,500 firms.
Search across state and federal law, get the citation, click through to the source paragraph.
Paxton extracts the line items and totals, then writes around them.
Identify risky clauses, missing indemnities, and inconsistencies across many documents at once.
Upload a 412-page deposition. Get a grounded answer, with the passage that proves it.
Harvey, Thomson Reuters' CoCounsel, LexisNexis Lexis+ AI, vLex Vincent, and Spellbook are all chasing the same lawyer. Most of them target large firms first. Paxton's edge is twofold: a benchmark it is willing to publish, and a price point that does not require a procurement department to clear.
It is a useful reminder that in legal tech, the largest opportunity is often the smallest customer. There are roughly 50,000 law firms in the United States. Twenty of them are big. The rest are who Paxton was built for.
But the lawyer is not reading them this time.
She has dropped the file into Paxton, asked her three questions, gotten her three answers with citations, and is now doing the thing the bar exam actually tested for: judgment. Reading the cases Paxton surfaced. Pressure-testing the argument. Drafting the parts that require her name on them.
That is the quiet reframe of what Paxton is doing. The point is not to replace lawyers. It is to take the long flat middle of a billable week - the reading, the cross-referencing, the first-pass drafting - and compress it into the part of the morning when the coffee is still hot. What's left is the work lawyers actually trained for.
Paxton's CEO will tell you, in plain language, that the goal is speed without sacrificing accuracy. The benchmarks suggest the company is closer to that than most of its competitors are willing to be measured on. The customer count suggests the rest of the legal profession is starting to notice.
The brief still gets filed Friday. It just takes a Tuesday afternoon instead of a Tuesday week.