An AI Audit for Your Business: What It Covers and Costs

An AI audit for a business is an inventory plus an evidence check: every AI system in use, who owns it, what data it touches, what it is allowed to decide, and whether any of that is written down. IBM found that 92 percent of organizations hit by an AI-related breach in 2026 had no AI access controls in place. The inventory is usually where the surprise is.

I am Amelia S. Gagne, CEO of Kief Studio. This page is what an AI audit covers and what it costs. The hands-on version is how to audit what AI is actually doing, and the policy that should exist first is how to build an AI policy for your team.

What does an AI audit for a business actually cover?

Six areas. They are the same six whether the auditor is external, internal, or you with a spreadsheet.

1. Inventory. Every AI system in use, including the ones nobody approved. Standalone tools, AI features switched on inside existing SaaS, API integrations, browser extensions, and anything an agent runs on a schedule. The features inside tools you already pay for are the most-missed category.

2. Data exposure. What each system reads, what it retains, where it sends it, and whether your material trains anyone's model. This is a vendor-contract question as much as a technical one, and it overlaps with what to ask vendors about protecting proprietary data.

3. Access and identity. What credentials each system holds and what those credentials can reach. An assistant with a shared human login is indistinguishable from the human in your logs, which is the case for non-human identity.

4. Decision authority. Where output is advisory and where it acts. Drafting an email is not approving a refund. The sequence problem, where individually permitted actions add up to something nobody authorized, is checking the sequence, not each click.

5. Accuracy and oversight. How output is checked, by whom, how often, and what the error rate actually is. Most organizations cannot answer the last part, which is itself the finding.

6. Documentation. Policy, training records, vendor terms, and a decision log. This is what an external assessor or customer questionnaire actually asks for.

What does an AI audit cost?

Three tiers, with the honest caveat that "AI audit" is not a standardized product and the term covers very different things.

Internal review. Your own team, a week of effort, a spreadsheet and a set of interviews. Cost is staff time. For a company under a hundred people with no regulatory driver, this is usually the right first step, and it produces most of the value because the inventory is the value.

External assessment. A consultant produces a findings report against a framework, commonly the NIST AI Risk Management Framework. Scope drives price. This is a report, not a certificate.

Certification. ISO/IEC 42001, published December 2023, is the first international AI management system standard. Annex A defines 38 controls across ten clauses in the ISO Harmonized Structure, so it integrates with ISO 27001. Certification requires an accredited third-party audit, and 2026 market rates for a mid-market organization run roughly €8,000 to €18,000 for the combined Stage 1 and Stage 2 initial audit, plus annual surveillance audits around €4,000 to €9,000. Other practitioner estimates put total first-year cost nearer $20,000 to $60,000 once internal build and consulting are counted. Certificates last three years. An existing ISO 27001 certificate cannot be extended to cover AI; it is a separate audit, though integrated audits reduce cost.

One caveat worth knowing before anyone buys: as of 2026, ISO 42001 is not a harmonized standard under the EU AI Act, so it does not grant a presumption of conformity. It is strong evidence of governance, and it is increasingly named in EU public procurement and enterprise vendor qualification, which is the real reason most companies pursue it.

Who actually needs which tier?

Most companies asking about an AI audit need the first tier and have been quoted for the third.

Run the internal review if you have AI in use and no written policy. That is the majority position and the gap is the policy, not the audit.

Get an external assessment if AI output touches a decision that affects a person, such as hiring, credit, pricing, or clinical support, or if a customer has started sending you AI questions in security reviews.

Pursue ISO 42001 if you sell AI-enabled products into enterprise or public sector procurement, or if a contract names it. It is a sales asset before it is a safety asset, and there is no dishonour in that as long as the sequencing is deliberate.

The failure mode is buying tier three to answer a question tier one would have answered. That is recommendation before diagnosis, and AI governance is a fertile field for it right now.

How do you run the inventory yourself?

The part that produces the most value and costs the least.

1. Pull the expense report and the SaaS bill. Every AI subscription someone expensed, and every existing tool that has shipped AI features since you bought it. The second list is longer.
2. Ask the team directly, without consequences attached. People use what helps them. If asking is risky, you get a clean inventory of nothing. Shadow AI is governed by enabling, not by a ban list.
3. Check OAuth grants and connected apps in your identity provider, Google Workspace, or Microsoft 365. Anything authorized against your data appears here.
4. For each entry, write five fields: owner, data it touches, whether it can act or only suggest, vendor retention and training terms, and who checks the output.
5. Sort by decision authority, not by popularity. The tool that can send, pay, or delete outranks the one that drafts.
6. Write the policy from what you found, not from a template. A policy describing systems you do not have is not a control.

A first pass takes a few days at a small company. The inventory is the deliverable. Everything downstream, including a certification scope, is built on it, and buying an audit without it means paying someone to do step one at consulting rates.

What findings come up most?

Four, near-universally.

Systems nobody owns. Introduced by someone who has since changed roles. Still connected, still authorized.

Credentials broader than the task. The integration was given full access because scoped access took longer to configure. Same finding as in mixed data environments, different context.

No accuracy baseline. Output is used but nobody has measured how often it is wrong, so there is no threshold at which anyone would stop.

Policy that does not match practice. A document written from a template, approved, and unread. The gap between it and the inventory is the real report.

None of those require an AI specialist to find. They require someone with the standing to ask and the patience to write it down. The same pattern shows up in the last mile problem in AI projects: the technical part is rarely what stalls.

Where this sits next to your security program

An AI audit is not separate from security. It is an access review with a new category of principal. The AI system holds credentials, reads data, and takes action, which makes it subject to the same questions as any other account.

Treating it as a new discipline is how organizations end up with an AI governance committee and no inventory. Treating it as access management extended to a new kind of actor is how it gets done. If you are earlier than this, figuring out where AI fits comes before auditing it, and data governance is the prerequisite for both.

At Kief Studio we run this as part of operating the stack rather than as a separate engagement, because the inventory is only current if someone maintains it. That model is at ltfi.ai. Our dependency and package tooling is public at kief.dev, including the Vekt lockfile scanner, and at ks-aur-scanner. Brian Gagne has handled the security architecture across the fourteen years we have worked together, at briansgagne.com.

Related reading

• How to audit what AI is actually doing
• How to build an AI policy for your team
• Shadow AI: govern by enabling
• What is a non-human identity for an AI agent
• Data governance is the prerequisite for AI

Frequently Asked Questions

What is an AI audit for a business?

It is an inventory of every AI system in use plus an evidence check on each one: owner, data touched, credentials held, decision authority, oversight, and documentation. The inventory is the core deliverable and usually the most surprising part.

How much does an AI audit cost?

An internal review costs staff time, typically a few days. An external assessment against the NIST AI Risk Management Framework produces a report and is scope-priced. ISO/IEC 42001 certification runs roughly €8,000 to €18,000 for the initial accredited audit, with practitioner estimates of $20,000 to $60,000 all-in for the first year.

Is ISO 42001 required by the EU AI Act?

No. As of 2026 it is not a harmonized standard under the Act and does not grant a presumption of conformity. It is voluntary, and it is increasingly named in EU public procurement and enterprise vendor qualification.

Do I need an AI audit if we only use ChatGPT and a few AI features?

You need the inventory. AI features switched on inside tools you already pay for are the most-missed category, and they are where data exposure usually sits. That is a spreadsheet exercise, not a paid engagement.

What is the most common finding?

AI systems with no named owner and credentials broader than their task. IBM reported that 92 percent of organizations with an AI-related breach in 2026 had no AI access controls in place.