An AI audit is a structured review of how your company actually uses AI — every tool in use, what data flows into them, what they cost, whether anyone checks their output, and who is accountable for the answers. Companies audit their finances annually and their security regularly. AI, which now touches customer data and customer-facing output at most businesses, usually gets neither.

The trigger is rarely curiosity. It's a board question nobody could answer, a vendor renewal nobody remembers approving, or the quiet realization that employees are using AI tools nobody sanctioned. Here is the framework I use in advisory engagements — usable internally, in an afternoon-to-a-week depending on your size.

The five audit questions

1. What is actually in use? (Inventory) List every AI tool touching company work: sanctioned platforms, features embedded in existing software (your CRM's AI assistant counts), and the shadow layer — personal-account tools employees use unofficially. Run an amnesty week: ask every team what they really use, with no penalties. The audit is worthless if people hide the truth, and the shadow inventory is usually 2–4× the official one.

2. What data does each tool touch? (Exposure) For every tool in the inventory: what classes of data enter it — public, internal, or restricted (customer PII, financials, contracts, credentials)? Under what account — company workspace or someone's personal login? Does the vendor train on your inputs? This is where audits find their scariest results, and the fix is usually a one-page policy plus account migrations, not a purge.

3. What does it all cost? (Spend) Sum the subscriptions — including seats on personal cards being expensed — then add the real number: time spent fighting, correcting, or duplicating AI work. Audits routinely surface three tools doing the same job and paid seats nobody's logged into for months.

4. Is anyone checking the output? (Quality) Where does AI-generated content reach customers, contracts, or decisions — and is there a named human reviewer at each of those points? Sample real outputs. Wrong-but-confident answers that shipped are the finding that turns audit into action.

5. Who is accountable? (Governance) If a tool leaked data tomorrow, who leads the response? If two vendors pitch overlapping products next week, who decides? If the answer is a shrug or a debate, the audit's top recommendation writes itself: name an owner.

Red flags that show up over and over

Restricted data in personal-account chatbots (the single most common finding)
AI features silently enabled in existing SaaS, processing customer data with no review
Paid tools nobody uses next to overloaded tools everybody uses
"The AI wrote it" offered as an explanation for an error that reached a customer
No written policy — or a policy nobody can locate

What to do with the findings

Rank findings by exposure, not by embarrassment: restricted-data leaks first, unreviewed customer-facing output second, waste third. Then fix the system rather than the symptoms — a policy people read, an owner people know, an approval path faster than the shadow path. The five-pillar governance framework is the standing structure an audit's findings should flow into, and the AI readiness assessment gives you a scored baseline to measure the next audit against.

An internal audit finds most of this. What it can't provide is independence — an internal owner grading their own governance, or a team auditing tools they chose. When the stakes justify outside eyes, an independent review is exactly the kind of engagement I take on through AI strategy & advisory, and a Strategy Intensive is the fastest way to pressure-test your worst finding.