Reconstruct the decision,
not just the model output.

A model response can be logged perfectly and still leave the organisation unable to answer the questions that matter: which evidence was available at the time, whether the premises were current, what counter-evidence was considered, who had authority, which action was actually executed and what happened afterwards.

An audit trail for consequential AI therefore has to join technical state with decision context and authority.

Reconstructability is strongest when the decision record is created as part of the workflow rather than assembled manually after an incident or review.

The record should be inspectable without implying that every data element must be duplicated into a new system. References, immutable identifiers, version binding and provenance can preserve the chain while source data remains in the systems that already own it.

For model validation, risk oversight, claims, underwriting, fraud intervention, agentic workflows or another consequential process, the useful question is not only “what did the model say?” It is “what decision was taken, under what evidence and authority, what action followed and what outcome later became observable?”

That chain creates a basis for validation, incident review, governance oversight and learning.

What is an AI audit trail?

A reconstructable record of the evidence, system state, proposed decision, checks, authority, action and outcome surrounding an AI-supported decision.

Why is logging model output not enough?

Because model output alone does not show which evidence was available, what assumptions were checked, who had authority, what downstream action occurred or what outcome followed.

What should an AI decision record contain?

Evidence provenance, relevant model or agent state, proposed decision, counter-evidence, evaluation, human approval or rejection, resulting action and observable outcome.