# Legal AI MOC Legal AI splits into two businesses: software sold to law firms, and AI-native law firms that sell the legal work itself. ## 1. The thesis: legal is a services market - Software is a $0.2T market. Services are $5.5T. See [[AI Eats Services Not Software]]. - Legal is a services market. Law firms bill hours, not seats. - An AI-native firm sells the outcome: a signed contract, a reviewed MSA. It does not sell a tool. - Price follows the output: per deal, per page, flat fee. See [[Outcome-Based Pricing]]. - The moat question from the services note applies directly: can the firm guarantee an outcome at a price below a human-only firm? ## 2. The two models | Model | Sells | Buyer | Examples | |---|---|---|---| | Legal AI software | Seats or usage | Law firms, in-house teams | [[Harvey]], [[Legora]] | | AI-native law firm | Legal work, flat or per-unit price | Companies that need legal work done | [[Crosby]], [[Moritz]] | | Open-source legal AI | Free software, bring your own model key | Law firms that self-host | [[Mike OSS]] | | Hybrid: firm and software company | Legal work and software | Companies and law firms | [[Omnilex]] | ## 3. AI-native law firms ### [[Crosby]] - New York law firm. Lawyers and software engineers on one team.^[https://techcrunch.com/2025/06/17/sequoia-backed-crosby-launches-a-new-kind-of-ai-powered-law-firm] - Reviews and negotiates commercial contracts (NDA, MSA) for startups. Cursor, Ramp and Cognition are customers.^[https://sequoiacap.com/article/partnering-with-crosby-a-law-firm-at-the-speed-of-ai] - Turnaround is about one hour. Price is per deal or per page, not per hour. - Founders: Ryan Daniels, John Sarihan. Sequoia-backed. Seed of $5.8M in June 2025.^[https://artificiallawyer.com/2025/06/19/crosby-raises-5-8m-seed-for-hybrid-ai-law-firm/] Later rounds are reported, but the sources are weak. Verify before you cite them. ### [[Moritz]] - San Francisco law firm with Norwegian founders: Pamir Ehsas, Stefan Mandaric. YC spring 2026 batch.^[https://www.ycombinator.com/companies/moritz] - Claim: the AI does 80% of the work. The client pays lawyers only for the last 20%. - Flat prices, same-day delivery. One YC customer got an MSA the same day for $750. - Raised $9M in May 2026. The round closed in four days.^[https://sifted.eu/articles/y-combinator-alum-moritz-raises-9m] - Does not sell to law firms. It wants to be the law firm. - Risk: Dealroom says the pitch shows the promise but not yet the proof.^[https://dealroom.co/news/129620-moritz-raises-9m-seed-for-ai-powered-corporate-law-firm/] ### [[Omnilex]] - Two businesses together. Omnilex Partners is the law firm. Omnilex AI is the software company built next to it.^[https://omnilex.law/] - Licensed attorneys handle every matter in the firm. Price is "on outcomes, not hours", with scope and fees agreed before work starts. - Offices in San Francisco, New York and Zurich. It covers all US states and Swiss regulation (FINMA, SECO, MiCA). - Practice areas: startup law, funds and transactions, regulation, contracts, data and AI, insurance, IP. It does not list litigation. - Omnilex AI started at ETH Zurich in 2024. It sells legal research and AI commentary on Swiss and German law. Its seed was $4.5M, led by Founderful.^[https://www.artificiallawyer.com/2025/11/11/omnilex-bags-4-5m-for-legal-research-ai-commentary/] The site says more than 150 enterprises use it. - Why it matters: - It runs both models at once. The software pays for the R&D, and the firm sells the outcome. - It works in both legal families: Swiss civil law and US common law. See §7. ### Litigation first - Crosby and Moritz start with the easiest work: high-volume, standard commercial contracts. - Some teams, for example in Jordan, start with the hardest work: litigation. - This is the opposite wedge. It is interesting because: - Litigation is the highest-value work in a firm. It has the most hours and the most at stake. - The work is adversarial and depends on deadlines. A mistake has a direct cost in court. - A team that does litigation well can move down to contracts. A contracts firm cannot easily move up to litigation. - Jordan is a civil law country. See §7. ## 4. Open source: [[Mike OSS]] - Free, self-hostable legal AI. Built by Will Chen, who was an associate at Latham & Watkins.^[https://www.artificiallawyer.com/2026/05/04/mike-the-open-source-legal-ai-platform-will-chen-interview/] - Does document review, contract drafting, tabular extraction and workflow automation. - AGPL-3.0 licence. Users bring their own model API key.^[https://legaltechnology.com/mike-oss-open-source-legal-ai-tool-changes-the-negotiation/] - More than 1,000 GitHub stars in 72 hours after the May 2026 launch. The repo is now `Open-Legal-Products/mike`. - It puts pressure on the software model. A free self-hosted tool sets a price floor for Harvey and Legora seats. - It does not compete with the AI-native firms. They sell the outcome and the lawyer's liability, not the tool. ## 5. Model layer - Legal AI products run on open-weight base models, often Chinese ones. See [[Chinese Frontier Labs]]. - Harvey post-trained its own model on Kimi K3. See [[The Cross-Pacific AI Chessboard]]. - Legal drafting is the standard example of a domain fine-tune. See [[Pretraining, Post-Training and Fine-Tuning]], [[Parameter-Efficient Fine-Tuning (PEFT)]]. - Over-training on legal text narrows a model. See [[Parameters]]. ## 6. Diligence lens - Legal-specific evals are a source of edge. See [[Evals]], [[Domain Experts as Eval Builders]]. - Trust means proving the AI work is correct. See [[AI Verification]]. - The main failure mode is a firm that stays a consultancy. Track lawyer hours per matter over time. See [[Consultancy-to-Platform Transition]], [[Deployment Velocity]]. - Full checklist: [[AI Agent Vertical SaaS DD MOC]]. ## 7. Global scale: two legal families Most of the world's legal systems belong to one of two families. A legal AI firm that gets both right can scale across many countries. See [[Common Law vs Civil Law]]. | Family | Main source of law | Where it is used | What the AI must do | |---|---|---|---| | Common law | Case law (precedent) and statute | UK and most of the Commonwealth: US, India, Australia, Canada, Singapore | Find and reason over past judgments | | Civil law | Written codes | Continental Europe, Latin America, most of the Middle East (Jordan included), China, Japan | Interpret and apply the code | - Crosby is in the US, a common law country. Moritz has Norwegian founders, a civil law country, and works in the US. - Omnilex already works in both families: Zurich (civil law) and the US (common law). - Many countries mix the two. Some also use religious law, for example Sharia courts for family matters in Jordan. - Scale question: does one product cover both families, or does each family need its own product? > [!question] Open questions > - Does the AI-native firm keep its margin when lawyer hours set the cost floor? > - Does Mike OSS make legal AI software a commodity, and push value to the firms that hold the liability? > - Which matters move to AI-native firms first: high-volume standard contracts, or complex work? > - Is litigation first a stronger wedge than contracts first? > - Is civil law easier for AI? The codes are shorter and fixed. Common law needs search over many past judgments. ---- tags: #boat #moc #verticalAI #AIstrategy links: [[AI Agent Vertical SaaS DD MOC]] [[AI Eats Services Not Software]] [[Selling AI MOC]]