Legal went first

Kimberly Tan's survey of where enterprises are actually adopting AI, published by a16z in April, contains a sentence that would have read as a joke at any point in the previous thirty years: legal was "surprisingly one of the first-mover industries in AI."

Legal was the market software went to die in. Long sales cycles, a buyer who bills by the hour and therefore has no obvious interest in doing the hour faster, and partnerships that make capital approval a committee sport. The received wisdom was not wrong. It has just stopped describing the present.

Why the objection stopped working

The piece's explanation is the most useful part of it, and it is an argument about fit rather than about enthusiasm.

Traditional enterprise software offered lawyers very little, because static workflow tools do not accelerate unstructured, nuanced work. A matter-management system moves a document from one state to another. It does not read the document. Practically all of the labour that makes legal work expensive sat outside what the software could touch, so what firms bought was filing cabinets with better search.

AI lands somewhere different. Parsing dense text, reasoning across a lot of it at once, summarising, drafting a response — that is not adjacent to the job. For large parts of a working week it is the job. The value proposition stopped needing a translation layer.

And in places it has gone past productivity into revenue. Tan's example is Eve, which sells to plaintiff-side firms, where processing more cases is not a cost saving but a larger book of business. That is a different sale from a per-seat licence, and it is the sale that gets a managing partner's attention.

The numbers, and their shelf life

The piece cites Harvey at roughly $200 million in annualised recurring revenue within three years of founding, and Eve at more than 450 customers and a $1 billion valuation.

Both were true. Neither is current, and the gap is four months.

Our own funding survey has Harvey at $350 million or more annualised as of August 2026, against 1,300-plus customers and 142,000 lawyers, on $1.17 billion raised at an $11 billion valuation. Eve we have at 1,000-plus plaintiff firms, and the growth note in our data is exactly the one that dates the a16z figure: 450 to 1,000 firms in six months. The number Tan quotes is the number Eve had when the piece was written.

This is not a criticism of the article. It is the thing the article is about. A market that restates its headline figures every quarter cannot be tracked by reading a quarterly.

What "the results are clear" is clear about

Two companies is where the honest reading gets harder, and it is worth saying what our own data does to that sentence.

Of the 96 AI-native legal companies in our survey, 14 have put a dollar revenue figure on the record. Thirty-seven disclose nothing at all — not revenue, not customers, not a growth rate. Harvey and Eve are legible because they chose to be legible, and companies choose that when the numbers are good.

So the results are clear about the winners. Whether they are clear about the category is a different question, and one that a survey of the visible cannot answer. It is the same reason our funding page ranks on capital raised rather than revenue: fourteen disclosures spanning four reporting years, mixing ARR with marketplace revenue, would be a precision we do not have.

A later a16z chart makes the same shape of point more starkly, and needs the same caution. Drawing on OpenAI's "What Frontier Firms are doing differently" from 12 August 2026, it shows weekly active enterprise Codex users by job title indexed to 1 February 2026, with Legal at 108x — comfortably the steepest line on the chart, ahead of sales at 41x and engineering at 5x. That is a real signal about direction. It is also indexed growth with no base disclosed, and 108x from almost nobody is a different fact from 108x from a lot of people. Engineering sits at 5x because engineering started at the top.

Two things worth holding alongside it

a16z is an investor in the market it is measuring. That is not a gotcha; it is the ordinary condition of most legal AI market data, including the parts we publish. We sell software here too, and the disclosure figures above are ours. Read all of it with the position in view.

Adoption is not competence. The months in which these adoption curves went vertical are the same months in which the Sixth Circuit sanctioned two lawyers $15,000 each over fabricated citations and a court in Oregon handed down roughly $110,000. Fast uptake and correct use are independent variables, and nobody sends you the memo when the second one moves.

The part the survey does not reach

Tan's frame is enterprise adoption, so the piece looks where the budgets are: firms and corporate legal departments. That is the right lens for its question and it is why the answer is optimistic.

Our product catalogue counts 273 products across the AI-native legal market. Nineteen are sold to someone handling a legal problem without a lawyer. The first-mover story is real, and it is almost entirely a story about people who already had counsel getting better counsel.

That is the gap we build into, and we would say that, so weigh it accordingly.

---

The a16z piece is in the river with the rest of the month. Figures attributed to a16z are as published in April 2026; figures attributed to us are from our own surveys, compiled August 2026, with sourcing and confidence flags per row. Nothing here is legal or investment advice.