AI audit
What is AI audit?
A structured review that checks whether an AI system does what it claims and meets the rules that apply to it: accuracy, bias, security, documentation, governance. It can be internal or run by an accredited third party, as certification against ISO/IEC 42001 requires. The point is evidence, not assurances, that the system behaves as intended.
Why it matters
An audit is only as good as the trail it has to follow, and most systems are not built to leave one. You cannot check whether a model was trained on appropriate data if nobody kept a record of what it was trained on. You cannot show a decision was fair if the inputs behind it were never logged. The value of an audit arrives before the audit does. Knowing one is coming forces the discipline that makes it possible: documenting choices as they are made, keeping logs, noting who changed what. Organisations that treat an audit as a one-off scramble at the end usually find the evidence they need was never captured, and it cannot be recreated after the fact.
In practice
A team sets out to review a model in use and hits a wall: no record of the training data, no log of the decisions it made, no note of who adjusted it and when. The review stalls, not because the model is bad but because its history was never written down. The lesson lands the hard way. The record has to be kept from the first day, because it cannot be rebuilt later.
