Shadow AI Self-Audit: Find Out What Your Team Is Already Using
How do I find out what AI tools my employees are using?
There is more AI in your organisation than leadership thinks, and the reason is never defiance. It is that the approved path was missing, or slower than the deadline. Every check below is designed to find systems rather than culprits, because an audit that feels like a hunt produces a quieter organisation instead of a safer one.
Seven checks, no product required, about an afternoon for the technical half. Steps six and seven produce more than the other five combined and are the ones most audits skip.
1. Read the expense reports
How. Search the last six months of card statements and reimbursements for AI subscriptions. Look for the twenty dollar monthly charges rather than the enterprise ones, and search partial merchant names, because these appear under parent companies and payment processors rather than product names.
What you are looking for. Individually expensed AI subscriptions are the clearest possible signal: someone needed a tool badly enough to pay for it and to ask for the money back. That is not a discipline problem, it is a procurement failure with a receipt attached.
2. Check your identity provider for third party sign-ins
How. In your SSO or identity provider, list applications users have authorised with their work account, including ones nobody approved. Most providers keep this and almost nobody reads it.
What you are looking for. Every entry is a service holding a token to something of yours. Pay attention to anything granted access to mail, files or calendars, which is a much larger exposure than a chat box and is usually invisible in any other check.
3. Look at outbound network or DNS logs for one ordinary week
How. Ask whoever runs your network for the top destinations by request count for a normal week, filtered to known AI domains. If you have no logging, note that as a finding rather than a blocker.
What you are looking for. Volume and spread. Ten people using something occasionally is a different problem from two people using it constantly, and the second usually means a workflow now depends on it.
Personal devices and phones on mobile data will not appear here at all, so read this number as a floor and never as the total.
4. Look for the tells in the work itself
How. Not to accuse anyone. To find where AI is load bearing. Documents that appeared implausibly fast, code with an unfamiliar house style, sudden uniformity in tone across people who write differently.
What you are looking for. Where the output is already depended on. That is where an unreviewed error will do damage, and it is the part of the map worth having.
Do not use this to confront an individual. It is unreliable as evidence about a person and using it that way ends the audit, because everybody stops talking on the same day.
5. Inventory the AI already inside tools you bought
How. List your existing SaaS and check which have added AI features, then check whether they are on by default. Your CRM, help desk, note taker, recruiting system and document suite have almost certainly shipped something since you last looked.
What you are looking for. This is where most organisations are already processing customer data through a model without a decision being made anywhere. It usually dwarfs the individual chat accounts and it is invisible to every other check on this list.
6. Ask, anonymously, and mean it
How. One anonymous form, three questions. What AI tools do you use for work, what do you use them for, and what would have to be true for you to use an approved tool instead. No names, no manager routing, and say plainly that nobody is in trouble.
What you are looking for. The third answer is the whole point of the exercise. It tells you exactly why your sanctioned path is losing, and it is the only question here whose answer you can act on directly.
This produces more than every technical check combined, and only if people believe the promise. If your organisation cannot make that promise credibly, that is itself the most important finding of the audit.
7. Ask the people who ship fastest what they actually do
How. Take your three most productive people, individually, and ask them to walk you through how a real piece of work got done last week, step by step, tool by tool.
What you are looking for. The workflow that is already working. Most organisations discover their best practice was invented by one person and shared with nobody, and the fastest safe path forward is usually to sanction it rather than to design something new.
What to do with what you find
1. Publish what you found, without names
Counts and categories, shared with everyone. It tells the organisation this was about systems rather than a hunt, and it makes the next audit possible. An audit whose results disappear teaches people to answer the next one carefully.
2. Fix the reason before you write the rule
Every finding is a workflow somebody needed. Sanction a tool that does the job, make approval faster than the workaround, and the volume of shadow use falls without a single prohibition. Write the rule second, once you know what it has to permit.
3. Turn off what nobody chose
The AI features that switched themselves on inside tools you already pay for are the fastest thing on this list to fix, and the one with the most customer data behind it. Decide each one deliberately, then record the decision.
4. Write the short list of what never goes in
Six specific categories beats a general prohibition, because a person can check themselves against a list and cannot check themselves against a mood. Our Paste Test has the categories and our policy builder writes them into a document.
Questions people ask
What is shadow AI?
Employees using AI tools the organisation has not approved, usually on personal accounts, to get work done. It is rarely defiance. It happens when the sanctioned path is missing, slower than the deadline, or forbids the work people actually have to do, so the practical fix is a good approved tool and a fast approval route rather than a stricter ban.
How do I find out what AI tools my employees are using?
Seven checks, none of which needs a product: expense reports for individually paid subscriptions, your identity provider for third party sign-ins granted with work accounts, outbound network logs for one ordinary week, the work itself for where AI is already load bearing, your existing SaaS for AI features that switched on by default, an anonymous survey, and a walkthrough with your fastest people. The anonymous survey produces more than all the technical checks combined.
Is shadow AI a security risk?
The exposure is real and it is not mainly about the chat box. Third party applications authorised with work accounts hold tokens to mail, files and calendars, and AI features inside SaaS you already bought are often processing customer data by default with no decision recorded anywhere. Individual chat accounts matter most for what gets pasted into them, which is a policy and category problem rather than a tooling one.
Should we ban unapproved AI tools?
A ban moves the activity to personal devices and personal accounts, where you have no contract, no logs and no visibility, and converts a manageable risk into an invisible one. What works is naming an approved tool that genuinely does the job, answering approval requests faster than the deadline, forbidding a short and specific list of data categories, and making it safe to report a mistake the same day.
How long does a shadow AI audit take?
An afternoon for the technical checks if you have the logs, and about a week in total including the anonymous survey and the walkthroughs. The survey needs a few days open to get honest answers, and it is the step most worth waiting for.
The two that follow this
The Paste Test gives you the short list of categories that never go in, which is the rule your findings will need. The policy builder writes it into a document in about two minutes, with no email required.
Information, not legal advice, and nothing here is collected. Before running the technical checks, confirm what your own privacy notices and any works council or employee representation agreement permit, because monitoring has its own rules and this page cannot know yours.