How to create an evidence trail for AI decisions

When AI is used in legal work, the question that eventually arrives is: how did you get to that answer? Being able to answer it is a documentation problem more than a technical one.

Log the instruction. Record the exact criteria or prompt applied, and version it. If the criteria changed mid-review, the record should show when, and which documents were assessed under which version.

Log the output. Store the model's determination and any confidence signal alongside the document, rather than overwriting it with the final human call. Both matter later.

Keep the citation. The single most useful artefact is a pointer to the specific text the model relied on. A conclusion with a quoted source can be checked in seconds; one without it cannot be checked at all. This is also the most direct protection against a model asserting something the record does not support.

Record the human decision. Note who reviewed the suggestion, whether they accepted or overturned it, and when. Overturn rates are a useful quality signal in their own right.

Keep the validation record. Sampling results, measured recall against a control set, and any QC passes belong in the same file. This is what turns "we used AI" into "we used AI, and here is how we tested it."

Much of this mirrors long-standing practice for technology-assisted review, where courts have generally focused on the reasonableness and documentation of the process rather than on the specific technology used.

Related reading: Lessons from Zhang v. Chen: How Canadian Lawyers Can Actually Avoid AI Hallucinations in Review.

Claira is designed so that each of these artefacts is produced as a by-product of normal review rather than assembled afterwards. You can see what the record looks like.