The State of Legal AI, August 2026

For the last month we have been counting things. Who raised money in legal AI, what they built with it, who it is sold to, how it is measured, and what changed when Anthropic, Google and Microsoft started shipping legal products of their own.

The two datasets are already public: a funding survey of 95 AI-native companies and a product catalogue of 269 named products. This is the summary of what they show. The long-form report follows.

The ten things worth knowing

1. The money is concentrated and the middle is thin. $6.20 billion across 95 AI-native companies. The top five hold 46% of it. The median company has raised $21 million and 51 of the 95 are under $25 million.

2. Almost nobody will say what they earn. 14 of 95 disclose a dollar revenue figure. Another 23 give a growth rate with no base. 37 companies — 39% of the list, some worth over a billion dollars — disclose nothing at all.

3. One in five sells legal work, not software. Nineteen of the 95 deliver legal services directly, either as firms built that way (Crosby, Keith, Lawhive, Moritz) or as software companies that bolted a firm on the side (Norm Ai, Eudia, EvenUp, Aavalynx). Arizona's alternative business structure licence is the mechanism for most of the US ones.

4. The buyer is in-house, not the law firm. Of 269 products, 189 are sold to corporate legal departments and 161 to law firms. Independent lawyers can buy 38. People without a lawyer can buy 19, and most of those are services with an attorney attached rather than software.

5. Agents are the largest category, and the word means four things. Fifty products are agents or agent infrastructure — more than any other type. In practice that covers renamed chat assistants, multi-step workflow runners, builders for making your own, and a small number of genuinely autonomous systems.

6. The autonomous ones do unglamorous work. Chasing clients, answering routine questions, requesting records, placing calls. Not arguing motions. That is the right place to start and it is worth noticing that nobody credible is claiming otherwise.

7. The connector layer may matter more than any single product. Trellis, midpage, Twin1, Clerq and Courtroom5 all expose legal data to general-purpose assistants over MCP. If the corpus is callable from anywhere, the window a lawyer works in need not belong to a legal software vendor.

8. The model labs arrived, with five different strategies. Anthropic shipped 12 practice-area plugins and 20+ connectors in May. Google launched an enterprise legal product in August with Harvey and Legora in its own connector list. Microsoft put contract redlining inside Word, built largely by engineers it hired out of a collapsed legal AI startup. OpenAI hired a legal lead and shipped nothing. xAI has done nothing at all.

9. The benchmarks are mostly written by the people being tested. Four vendors built legal benchmarks and all four won their own. The most useful number in the field is the gap between scoring methods: the leading model scores 90.6% on partial credit and 55.3% when every required element must be present. Legal work is conjunctive.

10. "Hallucination-free" was not. Stanford's RegLab measured Lexis+ AI hallucinating on 17% of queries and Westlaw's AI-Assisted Research on 33%, both while being marketed on the absence of that problem. Retrieval reduces hallucination. It does not remove it.

The one we did not expect

Our catalogue starts from a funding survey, which means it can only see companies that raised enough money to be reported. That filter has a blind spot and we walked into it.

Courtroom5 has raised about $420,000 since 2017, mostly grants. It is also the most substantial software product built for people representing themselves, it covers debt collection and foreclosure, and it appears in both Anthropic's and Google's connector lists alongside Harvey, Thomson Reuters and Relativity.

A method that follows the money cannot see the company that did not take much. That is worth stating plainly in a report built on funding data.

Why the agent question is the one that matters

Everything above is a snapshot. The part that will still be true in a year is the shift from tools that answer to systems that act.

Fifty products already claim it. Two of the three model labs with legal offerings ship agent builders. The connector layer is being laid now. And across all 269 products we found exactly one built to supervise other agents — Norm Ai's Supervisory AI — against fifty agents and counting.

We also found no published reliability numbers for any of them. Not from Harvey, not from Legora, not from Google. Task completion rates, intervention rates, error rates: nobody is saying. That absence is the single largest open question in legal AI right now, and it is the spine of the report.

What is in the report and not here

The blog posts behind this summary cover the funding picture, what got built, the market thesis, tools for people without lawyers, tools for solo practitioners, transcription accuracy, agents, benchmarks, and the model labs.

The report pulls those together and adds the work that does not fit a blog post: the full agent taxonomy with every product mapped to it, the regulatory picture across Arizona, Utah, England and Wales, jurisdiction and practice-area variance in model performance, the exits and failures of the last eighteen months, and what we think happens next — stated as predictions, so they can be checked against later.

Figures are as disclosed publicly and are not independently audited, and each is dated in the underlying pages. Nothing here is legal or investment advice.