Harvey’s two 2026 rounds show where legal AI is spending

Harvey’s $200 million March round valued the company at $11 billion. Its stated use of funds included expanding legal agents and the engineering teams that work inside customer organisations. That is a specific operating choice: selling a legal AI system also involves helping a firm adapt it to its documents, procedures and controls. Harvey reported more than 25,000 custom agents on its platform and more than 100,000 lawyers across 1,300 organisations. Those are company-reported usage and customer figures, not independent measures of legal accuracy. Primary announcement.

The $550 million round announced on 9 September came at a $15.5 billion valuation, co-led by Diffusion and Lightspeed. Harvey linked it to its first post-trained open-weight model and its Legal Agent Benchmark. It reported adoption by 80% of the Am Law 100, alongside corporate legal teams. The emphasis had expanded from implementing workflows to helping customers develop and control more of their own intelligence. A financing valuation records terms agreed by investors; it does not measure whether a particular agent can handle a client’s case reliably. Primary announcement.

Our reading is that the competition is moving beyond access to a general model. The implementation team, customer-specific workflows and evaluation process increasingly form part of the product. A legal department comparing vendors should ask for task-level evidence: which documents were used, who reviewed the result and what happens when an answer is wrong. A benchmark built by the vendor can make performance more visible, but its methodology and relevance still need inspection. Neither a larger round nor a larger customer count answers those questions. The practical test remains whether a firm can deploy the system inside its own professional responsibilities.

For a contract-review deployment, for example, ask the supplier to demonstrate a missed clause, a conflicting instruction and a document it is not allowed to read. The demonstration should include what the lawyer sees when the system cannot finish. Those ordinary failure cases reveal more about the service being bought than a general claim that agents perform complex work.