Ryan Walker was CTO of Casetext, the legal AI company Thomson Reuters acquired for $650 million in 2023. He then ran the CoCounsel product as a VP at Thomson Reuters. Javed Qadrud-Din built the first semantic search system in legal at Casetext — before GPT existed — and served as Head of AI, then Director of ML at Thomson Reuters. J.P. Mohler practiced at WilmerHale and Cooley and worked as a senior ML researcher at Thomson Reuters post-acquisition.

This is not a team that discovered law as a vertical. These are the people who trained the models that power the tools every BigLaw firm now uses. They know exactly where the AI is good, where it hallucinates, and how to design human-AI handoffs that do not embarrass anyone in a deposition.

Javed's early work at Casetext is particularly notable: he built autoencoder-based semantic search with auxiliary loss functions to compress legal document vectors without losing retrieval precision — before the modern transformer era made this table stakes. The intuition developed there — that legal text has structure that generic models miss — clearly informed how General Legal designs its contract review pipeline today.

How It Works (Technical Architecture)

General Legal has not published its full stack, but from what is publicly known, the architecture follows the pattern the team developed at Casetext and Thomson Reuters:

Document ingestion layer: Contracts arrive via Slack or email. The system parses and normalizes contract structure — identifying clause types, obligations, representations, and risk provisions using fine-tuned classifiers trained on commercial agreement corpora. Off-the-shelf LLMs struggle here because commercial contract language is dense with terms of art that require domain-specific training data.

Risk analysis and clause comparison: Each clause is scored against a playbook of acceptable terms. The playbook is likely per-client, built from their historical contracts and preferences. This is the "AI doing the actual work" layer — flagging non-standard indemnification caps, missing limitation-of-liability provisions, aggressive IP assignment language.

Attorney review layer: Here is where General Legal's model diverges from pure automation. Human attorneys with actual law licenses review the AI output before anything goes back to the client. This is not window dressing — it is the UPL (unauthorized practice of law) compliance mechanism, and it is what allows them to operate as a law firm rather than a software vendor. It also means the AI is doing 80% of the work and the attorney is doing error correction and judgment calls. That is a 5x-10x leverage ratio per attorney.

Negotiation tracking: Redlines go back via Slack. The system likely tracks version history and outstanding issues, giving attorneys a clear view of what has been accepted, rejected, and is still live. Think contract lifecycle management baked into the workflow rather than bolted on as a separate SaaS layer.

Difficulty Score

Overall: 7/10. The technical moat is real but the harder moat is institutional — law firm licensure, attorney hiring pipelines, and malpractice insurance. This is one of the few AI startups where the non-technical barriers are higher than the technical ones.

The Moat

Hard to replicate:

• Law firm licensure. Getting licensed to practice law in California (or any jurisdiction) takes months and requires real attorneys. You cannot vibe-code your way to a bar license. This is a regulatory moat most tech startups cannot navigate fast.
• Training data. The team's access to Casetext's proprietary legal corpus and Thomson Reuters' contract data is not available to anyone starting from scratch. Their models know things generic GPT does not know about how commercial lawyers actually negotiate.
• Team credibility. Growth-stage GCs and CFOs are not going to hand their MSAs to a team with no legal credentials. The CoCounsel pedigree opens doors that no amount of fundraising can substitute for.
• Attorney trust network. Law is a relationship business. Once you are embedded as the commercial contracting firm for a Series A company, you grow with them. Switching costs are real — they have all your playbooks, your redline history, your negotiating positions.

Relatively easy to replicate:

• The Slack integration. Trivial to build.
• The flat-fee pricing model. Anyone can offer flat fees — but making money on flat fees requires the AI leverage ratio they have built.
• The contract types they focus on. MSAs and NDAs are not exotic — competitors can target the same paper.

The most interesting thing about General Legal's moat is that it gets stronger as they do more work. Every contract they review trains their playbook. Every client relationship generates more structured data about how companies negotiate. Over time, they will know more about what the market rate for an indemnification cap looks like in SaaS B2B contracts than any other entity on Earth. That is a compound moat, not a static one.

The Counterargument

The bear case on General Legal is the same as every vertical AI services company: what happens when the underlying AI gets 10x better? If GPT-7 can review contracts as well as a trained attorney, the value of the human review layer compresses. At that point, General Legal is competing against any competent developer who can wrap an API in a Slack bot.

The bull case is that this is exactly backwards. Better AI means cheaper operations, not commoditization — and the regulatory moat (you need a law firm to give legal advice, period) holds regardless of model quality. The team that is already operating at scale with better AI than anyone else will absorb the capability improvements and pass the savings on, widening the price gap against traditional firms rather than narrowing it.

The founders clearly believe the bull case. They have seen this movie before — they built CoCounsel when people said AI could not do legal research, and Thomson Reuters paid $650 million for the outcome.

Why This Matters Beyond Legal

General Legal is a template for a class of startup that is becoming possible for the first time: the AI-native professional services firm. Not a tool that sells to professionals. A firm that uses AI to deliver professional services directly, at a fraction of the cost, with meaningfully faster turnaround.

The same model is coming for accounting (tax prep, audit sampling), HR compliance (handbook review, offer letter generation), and financial advisory (plan document review, regulatory filings). Every professional services category with standardized deliverables, high incumbent pricing, and conservative adoption of technology is vulnerable to exactly what General Legal is doing.

General Legal is not interesting because it is building clever legal software. It is interesting because it is showing what it looks like when the people who built the tools decide to use the tools to run the firm themselves.

That is a different kind of startup. Pay attention.