As AI moves from the margins of legal work into the middle of it, one question keeps arriving from different directions: what happened. Which system was used, on what material, by whom, producing what, and who approved it. Insurers are beginning to ask at renewal. Clients are beginning to ask in outside-counsel guidelines. Opponents will ask eventually, and they will ask in a form that has to be answered.

Telemetry and evidence are different products

Most software logging exists to help the vendor: to debug an incident, to measure adoption, to bill. It is sampled, retained for a while, keyed on sessions rather than matters, and owned by the company that wrote it.

A legal audit trail exists to be produced. It has to be complete rather than sampled, organised around the matter rather than the session, retained on the matter's own schedule, and available to the firm without asking anyone's permission. Building the first and calling it the second is the most common failure in this category, and it is invisible until the day it matters.

What the record has to capture

Six elements, and the fourth is the one that is usually missing.

  • Who initiated the interaction, and under what role
  • Which client and matter it was bound to
  • Which model and version served it, and where that inference ran
  • What context was actually supplied — the retrieved documents and the assembled prompt, not a description of them
  • What was produced, and every human modification before it was used
  • Who approved it, against which version, and when

Without the fourth, the trail records that something was generated but not what it was generated from — which is precisely the question in any dispute about whether an output was reasonable to rely on.

What most platforms miss

Application-level logging is standard: queries submitted, documents opened, buttons pressed. The context supplied to the model usually is not logged, for understandable engineering reasons — it is large, it duplicates material already stored elsewhere, and nobody asked for it during the build.

The second common gap is the approval. Many systems record that a document reached a final state without recording who put it there, on which version, having seen what. A record of outcomes with no record of decisions cannot support a supervision argument under Rules 5.1 and 5.3, which is the argument it will be asked to support.

If you cannot produce a complete audit trail of every AI interaction associated with a matter, you cannot fully defend your work product.

Immutability and custody

A log that anyone can edit — including the vendor, including in good faith during an incident — is not evidence. The properties that make it evidence are ordinary and testable: append-only, cryptographically signed so alteration is detectable, held on infrastructure the firm controls, and queryable without the vendor being in the loop.

Custody is the part firms discover late. If producing the record requires a support ticket, then in the scenario where you most need it — a dispute with that vendor, an insolvency, a contract that has ended — you may not have it at all.

The test

There is a single question that settles whether a platform's logging is adequate. Pick a document that went to a client six months ago and ask the firm to reconstruct, without contacting the vendor, exactly what the model saw and who released it.

Most platforms produce a partial answer. The gap between that partial answer and a complete one is the risk a firm is carrying, and it is worth measuring before somebody else measures it for you.