Preparing an AI system for audit: the evidence you need before they ask
Regulators in the UAE and Saudi Arabia increasingly ask how AI systems decide. The seven artefacts that answer an audit calmly — and how to collect them as a by-product of normal operation.
An AI audit rarely announces itself as one. It arrives as a line in a central bank examination plan, a question from internal audit after an incident, a data protection office inquiry, or a national programme office asking how your flagship system makes decisions. Whatever the trigger, the question underneath is always the same: can you show — not tell — how this system works, and who is accountable for it?
Organisations discover the gap at the worst moment. The system works; the evidence does not exist. Nobody can produce the list of AI systems in production, let alone the validation results for the one under scrutiny. What follows is weeks of archaeology that a small amount of foresight would have made unnecessary.
The seven artefacts an audit actually asks for
Across sectors and regulators, AI scrutiny converges on a short list of evidence. If you can produce these seven, most examinations go smoothly:
- A system inventory. Every AI system in production, what it does, what data it touches, and — critically — a named accountable owner. "The vendor" is not an owner.
- Risk assessments, dated before deployment. Evidence that someone considered what could go wrong — bias, error impact, data exposure — and decided the mitigations, before go-live rather than after the incident.
- Validation results. How you knew the system was good enough to deploy: evaluation sets, accuracy on them, and acceptance thresholds. For Arabic-serving systems, that must include performance measured on Arabic — dialect included, not just formal text.
- Decision logs. For consequential decisions, a per-case record: the inputs, the model version, the output, and any human action taken on it. This is the artefact that most often does not exist.
- Human oversight records. Where the design says "a person reviews flagged cases", proof that reviews actually happen — approval timestamps, override rates, escalation outcomes.
- Change history. Models change: retraining, prompt updates, version migrations. Auditors ask what changed, when, why, and what revalidation followed. A system that changed silently is a finding in itself.
- Incident records. What went wrong, what you did, what you fixed. A short, honest incident log builds more trust than a claim that nothing has ever failed — which no examiner believes.
Evidence as a by-product, not a project
The expensive way to pass an audit is to reconstruct evidence when asked. The economical way is to design the system so evidence accumulates as it runs: logging built into the serving path, evaluation runs scheduled and archived automatically, risk assessment embedded as a gate in the delivery process, ownership recorded in the inventory when a system ships.
Done this way, audit preparation stops being an event. The answer to "show us" becomes an export, and the review meeting becomes short. This is the philosophy behind our governance work: the binder writes itself, or it will never be written.
Agentic systems raise the bar
The move toward agents — systems that take actions, not just produce text — raises audit expectations sharply, and both the UAE's federal agentic AI programme and sector regulators are alert to it. For agents, the evidence list grows: the exact permission boundaries each agent holds, logs of every action against those boundaries, the approval trail for consequential steps, and adversarial test results showing the boundaries hold under attack — in both Arabic and English, because attacks come in both.
If your roadmap includes agents, build this instrumentation from day one. Retrofitting an audit trail onto an agent already in production is the hardest version of the problem.
Start with one honest hour
You do not need a transformation programme to begin. List your AI systems in production. For each, ask: who owns it, how do we know it works, and what record exists of what it decided last month? Wherever the answer is "we're not sure", you have found next quarter's priority — before an examiner finds it for you.
Inspection-readiness compounds quietly: teams that can answer for their systems ship faster, negotiate procurement from strength, and take the regulator's call without dread. In a region whose national AI strategies explicitly pair ambition with responsibility, that is not overhead. It is the licence to keep building.