Can a firm use client work to improve its own AI tools?

Start with the clause, not the technique. Professional engagement terms typically say three separate things that bear on this: that the firm will keep the client's information confidential, that it will use that information only for the purposes of the engagement, and that some or all of the deliverable belongs to the client. The second is the one that catches internal tool building, and it is the one people skip — a purpose limitation is breached by using the material for a different purpose whether or not anybody outside the firm ever sees it.

Ownership is the second constraint and it varies more than firms remember. Where the deliverable is assigned to the client, the firm holds a licence to what it produced rather than the thing itself, and that licence is usually narrow. Where the firm retains ownership and grants the client a licence, the position is much better. Firms with a mixed history of engagement terms — most firms — cannot answer the question in general and have to answer it per client, which is itself a useful finding.

The distinction that firms reach for is between the client's information and the firm's own know-how, and it is a real distinction that does not do as much work as they hope. Nobody seriously argues that a lawyer forgets how to structure an argument between matters. What is contested is the artefact: a repository of past deliverables, an index over client documents, a model tuned on them. Those embed particulars rather than technique, and a search over them returns the particulars. The test worth applying is whether the tool can produce something a person could identify as coming from a specific client's matter.

Which suggests where the safe ground actually is, and it is substantial. Templates and precedents with the particulars removed by a person. A style guide derived from the firm's own writing. Checklists, process documents, and structured knowledge about how the firm works. A retrieval system over the firm's own published material and its internal guidance. All of these are the firm's own product and none of them requires reading anybody's engagement letter.

There is a separate question about the vendor, and it is often the one that decides the matter. Building an internal tool usually means sending material to somebody — a hosting provider, a model provider, an embedding service — which is a disclosure to a third party even when the tool is described as internal. That disclosure is governed by the same confidentiality clause and often by an explicit subprocessor provision. A firm can be entirely right that the tool is internal and entirely wrong that no disclosure occurred.

The route that works, where the value justifies it, is consent obtained properly rather than a clever reading. Asking clients — particularly institutional ones, who deal with this constantly and have positions on it — is a conversation firms consistently expect to go worse than it does. Some say no, which is information. Some say yes with conditions that are easy to meet. And a firm that asked has a durable answer, where a firm that reasoned its way to a permissive interpretation has an argument it will have to make at the worst possible moment.