Most of what you'll read about an AI audit assumes you build models: it talks about training data, model drift and explainability tests. Most companies of 20 to 1,000 people don't build models. They use them, through a dozen chat tools, assistants built into software they already pay for, and a few agents someone connected last quarter. For those companies, the hard part of an AI audit isn't testing a model. It's finding out what's in use at all.
What is an AI audit?
An AI audit is a structured check of the AI an organization uses: which tools, features and agents are in use, what data they can reach, what the vendors' terms allow, which decisions they influence, how their output is checked, and who owns each one. For a company that uses AI rather than builds it, the audit is mostly about use and data, not model code.
This guide is for the COO, IT lead, risk lead or HR lead who has been asked to "audit our AI" and wants to know what that means in practice. It covers internal audits of AI use, the bias audits some hiring laws require, and how certification fits in. It isn't legal advice.
A naming note. "AI audit" also means using AI to carry out a financial audit, and a lot of what ranks for the term is about that. This article is about auditing AI, not auditing with it.
Is an AI audit required by law?
In the US, there's no general legal duty to audit the AI you use, but some laws require a specific audit, and more ask for the records an audit produces.
| Rule | What it asks for | Who it reaches |
|---|---|---|
| NYC Local Law 144 | A bias audit by an independent auditor within one year before using an automated employment decision tool, a public summary of the results, and notice to candidates 10 business days before use | Employers and employment agencies using such a tool in New York City, enforced since July 5, 2023 |
| Colorado SB 26-189 | Records "necessary to demonstrate compliance with the act" kept for at least 3 years | Developers and deployers of tools that influence consequential decisions, including hiring |
| NIST AI Risk Management Framework | Nothing. It's voluntary guidance | Anyone who wants a structure to work from |
| ISO/IEC 42001 | Nothing, unless you choose certification. Then an accredited certification body audits your AI management system | Organizations whose customers ask for a certificate |
New York City's rule is the clearest case of a legally required AI audit. The city's FAQ calls a bias audit "an impartial evaluation by an independent auditor" and says it must at least calculate selection or scoring rates and impact ratios across sex, race and ethnicity, and intersectional categories. The auditor can't work for you or for the vendor, and can't have been involved in using or developing the tool. The vendor can commission the audit, but the city is clear that the employer remains "ultimately responsible" for making sure one was done. Our guide to the Colorado AI law covers what Colorado asks of employers from January 2027.
Outside those rules, the reason to audit is practical. You can't write a usable AI usage policy, answer a customer's security questionnaire or respond to a regulator's question about a tool you don't know you have.
What does an AI audit cover?
A useful AI audit for a company that uses AI covers six areas: the inventory, the data, the vendor terms, decisions about people, how output is checked, and who owns each tool. The inventory comes first because everything else is done row by row against it.
The inventory
List every place AI is used for work: the chat tools staff use (approved or not), the AI features inside software you already pay for, custom assistants and agents, and anything built on a model through an API. The second group is the one audits miss. Your applicant tracking system, CRM, help desk and meeting software may have added AI features that nobody bought as AI.
The first group is bigger than most leaders expect. In a June 2026 PagerDuty survey of 1,250 office professionals at companies with $500 million or more in annual revenue, conducted by Wakefield Research, 66% said they had used AI tools at work that they believed weren't permitted. 86% said their organization had an AI policy.
The NIST AI Risk Management Framework makes the inventory an explicit outcome: "Mechanisms are in place to inventory AI systems and are resourced according to organizational risk priorities" (GOVERN 1.6). Our article on shadow AI covers how to get people to tell you what they use without a ban that drives it underground.
The data
For each row, ask what the tool can reach and what people put into it. In the same PagerDuty survey, 88% said they had shared work-related information with public AI tools, and 31% had shared financial information or confidential company documents and strategies. Map the data classes (client data, personal data, health data, financial records, source code) against the tools allowed to touch them. The data map, in module 2 of our free course, is a one-page way to do it.
The vendor terms
For each vendor, record four things from its current documentation, with the date you checked: whether your data is used for training by default, how long data is kept, whether you can sign the contract terms you need (a data processing agreement, or a business associate agreement for health data), and what admin and audit controls your plan includes. Plan tier matters. Our comparison of ChatGPT Plus and Business shows how two plans at the same price differ on training defaults and ownership, and our guide to HIPAA compliant AI shows which plans sign a BAA.
NIST's framework treats this as its own outcome: policies that "address AI risks associated with third-party entities" (GOVERN 6.1).
Decisions about people
Flag every use that scores, ranks, recommends or classifies people: candidates, employees, customers, patients, tenants. Those are the uses hiring and consumer laws reach, and the ones where a bias audit may be required. A recruiter pasting résumés into a chatbot and asking who to shortlist counts, even though nobody bought the chatbot as a hiring tool.
How output is checked
For each use, write down who checks the output before anyone relies on it, and whether they have the time, the information and the authority to reject it. "A person reviews it" isn't an answer until you know which person, and what they look at.
Who owns it
Every row needs an owner, a backup, a way to switch the tool off, and a date to look again. NIST asks for documented "roles and responsibilities" (GOVERN 2.1) and mechanisms "to supersede, disengage, or deactivate AI systems" that behave outside their intended use (MANAGE 2.4). In a small company that's one name per row and one tested off switch.
Which AI audit framework should you use?
For a company that uses AI rather than builds it, the NIST AI Risk Management Framework is the most practical place to start: it's free, voluntary and organized around four functions you can audit against.
| Framework | What it is | When to use it |
|---|---|---|
| NIST AI RMF 1.0 (January 2023) | Voluntary US guidance with four functions: GOVERN, MAP, MEASURE and MANAGE | As the structure for an internal audit, at any size |
| ISO/IEC 42001:2023 | An international standard for an AI management system that organizations can be certified against | When customers or regulators ask for a certificate |
| IIA AI Auditing Framework | The Institute of Internal Auditors' framework for internal audit teams | When you have an internal audit function |
| NYC bias audit rules | A defined calculation of selection rates and impact ratios, done by an independent auditor | When you use an automated hiring tool in New York City |
NIST says its actions "do not constitute a checklist, nor are they necessarily an ordered set of steps." That's fair warning: the framework tells you what good looks like, not what to do on Monday. The checklist below turns it into questions you can answer for each tool.
ISO/IEC 42001 is a different kind of thing. UKAS, the UK's accreditation body, calls it "the world's first management system standard for artificial intelligence", and certification against it is done by accredited certification bodies. The first UK accreditation for that work came in January 2026, when UKAS accredited BSI. For most companies of 20 to 1,000 people it's worth knowing about and rarely the place to start. Get the inventory and the owners right first. Any management system, certified or not, is built on them.
AI audit checklist
Copy this checklist and answer it for each row of your inventory. Some questions gather facts and some test a control, so the finding isn't always a "no". Any "don't know" is a finding. So is any answer that shows a risk without a control: data your policy bars still reaching the tool, training switched on by default, a tool that ranks people with no bias audit or no named checker, no owner or no off switch. A finding isn't a failure. Write it down with an owner and a date.
Inventory
- Is this tool, feature or agent on the list, with its vendor, plan and the teams that use it?
- Have you checked the AI features inside software you already pay for, not only the tools bought as AI?
- Do you know whether staff use personal accounts for it?
Data
- Which classes of data can it reach: client, personal, health, financial, source code?
- Is any of that data barred from it under your policy or a contract?
- Can it connect to other systems (email, files, CRM) and act in them?
Vendor terms (record the date you checked)
- Is your data used for training by default on your plan?
- How long is data kept, and can you change it?
- Do you have the contract you need: a data processing agreement, and a business associate agreement if health data is involved?
- Does your plan include audit logs you can actually read?
Decisions about people
- Does it score, rank, recommend or classify candidates, employees or customers?
- If it's used in hiring in New York City, is there a bias audit from the past year, with a public summary and notices sent?
- Have you told the people affected, where a law requires it?
Output checks
- Who checks its output before anyone relies on it?
- Do they have the time, the information and the authority to reject it?
- Is there a record of the check for high-stakes uses?
Ownership
- Does it have a named owner and a backup?
- Can you switch it off, and has anyone tested that?
- Is there a date to review it again?
An audit is a list with an owner on every row
The most common failure in an AI audit isn't a missed risk. It's a report that lists risks and changes nothing, because nobody owns any of them. An audit that ends in a slide deck goes stale the week a vendor ships a new feature.
Our framework treats the audit's output as a working list, not a report: every tool on it has an owner, a backup, an off switch and a review date, and the list is kept up to date as tools come and go. Your AI list, in module 3 of our free course, builds that list. Owners and backups, in module 7, fills in the ownership columns. If the audit is the first step in deploying AI across the company rather than a one-off check, that list is what the rest of the rollout runs on.
How to run an AI audit
There are six steps, in order. Each one produces what the next one needs. For a company of 20 to 1,000 people, a first pass usually fits in about a month if one person owns it.
01 · Name an owner and set the scope
Pick one person to run the audit and give them a deadline. Decide the scope in a sentence: every AI tool, feature and agent used for company work, on company or personal accounts. Tell staff the audit is happening and that the point is to find tools, not punish people.
02 · Build the inventory from three sources
Ask people what they use, with an amnesty for anything unapproved. Pull what IT can see: single sign-on logs, expense reports, browser extensions and the admin consoles of the tools you already pay for. Then go through each major vendor's release notes for AI features switched on by default. Merge the three into one list. The free Shadow AI Audit gives you a starting list in about five minutes.
03 · Trace the data for each row
For each tool, record what data it can reach and what people put into it. Use the platforms' own logs where you have them. Microsoft, for example, says that if your organization has auditing turned on, Copilot interactions are logged automatically under Audit (Standard), including references to the files, sites and other resources Copilot accessed to answer a prompt. Logs for AI apps from other vendors are billed pay-as-you-go and kept for 180 days.
04 · Read the vendor terms and flag the people decisions
For each vendor, record the training default, retention, contract terms and admin controls on your plan, with the date you checked. Then mark every row that scores, ranks or decides about people. Those rows get the closest look, and any automated hiring tool used in New York City needs its bias audit checked.
05 · Record the findings against the checklist
Answer the checklist for each row. Sort the findings into three groups: stop now (data going where it mustn't), fix within 30 days, and accept with a reason. Keep, replace or retire, in module 3 of our course, helps you decide what to do with each tool.
06 · Give every row an owner and a review date
Close the audit by assigning each tool an owner, a backup, a tested off switch and a review date. Put the list where people can see it. The next audit is then an update to the list, not a fresh hunt. If you'd like a second pair of hands, our consultation is an audit plus a plan built on our framework, and AI governance covers ongoing support.
How much does an AI audit cost?
An internal audit of the AI you use costs mostly staff time: about a month of part-time work for one owner, plus a few hours from each team that uses AI. The tools to start are free, including NIST's framework and our Shadow AI Audit. Outside help costs more, and prices vary widely by scope. Our guide to AI consulting cost collects the fixed fees firms publish for assessments and audits. A bias audit of a hiring tool or an ISO/IEC 42001 certification is a separate engagement with an independent auditor or certification body.
What's still open
- More state laws. Colorado's law starts in January 2027, and other states are writing their own rules on AI in hiring and consumer decisions. The audit's inventory is what you'll need for each of them.
- Vendor features change monthly. An AI feature that's off today can arrive switched on next quarter. That's why the audit ends in a list with review dates, not a report.
- Certification demand. ISO/IEC 42001 certification is new, and it's not yet clear how often mid-size companies' customers will ask for it.
We'll update this article when the rules change. The date at the top of the page shows when we last checked.
Sources
- Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST, January 2023: the four functions, its voluntary status, the note that its actions aren't a checklist, and subcategories GOVERN 1.6, GOVERN 2.1, GOVERN 6.1 and MANAGE 2.4.
- Automated Employment Decision Tools (AEDT), NYC Department of Consumer and Worker Protection, checked November 2026: the bias audit, public summary and notice requirements, and the July 5, 2023, enforcement date.
- AEDT Frequently Asked Questions, NYC Department of Consumer and Worker Protection, June 2023: what a bias audit must calculate, who counts as an independent auditor, and the employer's responsibility.
- SB 26-189, Automated Decision-Making Technology, Colorado General Assembly, May 2026: the three-year record-keeping duty.
- Audit logs for Copilot and AI applications, Microsoft Learn, August 2026: what Copilot audit records contain, and the billing and 180-day retention for other AI apps.
- PagerDuty Report Finds Two-Thirds (66%) of Office Professionals Have Used Unauthorized AI Tools at Work, PagerDuty, June 2026: the survey of 1,250 office professionals by Wakefield Research, used for the figures on unpermitted use, data sharing and policies.
- UKAS grants first AIMS accreditation, UKAS, January 2026: the first UK accreditation of a certification body for ISO/IEC 42001:2023.
- The IIA's Artificial Intelligence Auditing Framework, The Institute of Internal Auditors, checked November 2026: the framework for internal audit teams.