Ask the leadership team of a mid-size company whether employees use AI at work, and the answer is usually “a little, maybe.” Ask the employees, anonymously, and the answer is different. Sales reps paste call notes into a chatbot to draft follow-ups. Finance pastes a spreadsheet into an AI tool to clean it up. Someone in HR asks an assistant to rewrite a sensitive performance review. A developer installs a browser extension that reads every page they open.
None of this was approved, and most of it was done with good intentions. That is shadow AI: AI tools used for work without the knowledge or approval of the people responsible for data, security and compliance.
Shadow AI is not a reason to panic, and banning every AI tool rarely works. People who found a tool that saves them an hour a day will keep using it, just less visibly. The goal is to see what is happening, decide what is acceptable, and give people a safe way to keep the benefits.
Why it matters
The risk is not that employees use AI. The risk is what they put into it and what the tool is allowed to do with it.
- Data leaving your control. Free and consumer versions of AI tools often have terms that allow the provider to keep inputs, and sometimes to use them to improve their models. Customer records, contracts, source code and financials can end up outside the agreements you have with clients and regulators.
- Compliance obligations. If you handle health information, payment data or client confidences, using an unapproved tool can breach HIPAA, contractual commitments or professional rules, even when nothing “bad” happens afterward.
- Accounts nobody manages. Personal sign-ups on company email create accounts that IT cannot see, offboard or audit.
- Decisions nobody can explain. When an AI tool drafts a customer answer or scores a lead and nobody records that it was involved, you lose the ability to review or correct it later.
Step 1: Ask first, and make it safe to answer
The fastest source of truth is your own people. Run a short, anonymous survey with three questions:
- Which AI tools do you use for work, including free ones and browser extensions?
- What do you use them for?
- What kind of information do you put into them?
Say clearly that the goal is to keep the useful tools, not to punish anyone. If people fear consequences, they will under-report and you will be governing a fiction.
Step 2: Check what your systems already show
You do not need new software to get a first picture. Most companies already have the data:
- Identity provider and SSO logs. Look for “Sign in with Google” or “Sign in with Microsoft” grants to AI apps.
- Expense reports and corporate cards. Small monthly subscriptions to AI tools are easy to spot.
- Browser extension inventories. If you manage company browsers, list installed extensions and flag the ones with broad page-reading permissions.
- Network or DNS logs. Traffic to known AI services shows usage that does not involve an account.
- Email. Welcome and billing emails from AI vendors sent to company addresses.
Combine the survey and the logs into one simple inventory: tool, team, use case, type of data, account type (personal or company), and whether a contract exists.
Step 3: Sort the risk
Not every tool deserves the same attention. A simple three-level sort works for most mid-size companies:
| Level | What it looks like | What to do |
|---|---|---|
| Low | Public information only, such as drafting a social post or summarizing a public article | Allow, with general guidance |
| Medium | Internal but non-sensitive information, such as meeting notes or internal process docs | Allow only on a company account with data-retention and training controls turned off |
| High | Customer data, health or financial records, contracts, source code, credentials | Stop until there is an approved tool with a signed agreement |
The question that decides the level is almost always the same: what data goes in, and where does it go after that?
Step 4: Give people a better option than the one you take away
If you block a tool that a team relies on without offering an alternative, usage moves to personal phones and laptops, where you see even less. For each high-value use case you find, offer an approved route:
- A company account on a business or enterprise plan, where the vendor contractually commits not to train on your data.
- An internal assistant connected to your own documents, with access controls that match the ones you already use.
- Clear examples of what is fine to paste and what is not.
Step 5: Write it down in one page
Turn what you learned into a short acceptable-use policy: which tools are approved, what data is allowed in each, how to request a new tool, and who to ask. Keep it to one or two pages so people actually read it.
Step 6: Repeat every quarter
New AI tools appear every month, and existing tools add AI features without asking. A quarterly check of the same sources (survey, SSO grants, expenses, extensions) keeps the inventory honest without turning it into a full-time job.
Where frameworks fit
If you want a reference model, the NIST AI Risk Management Framework and its Generative AI Profile describe how to govern, map, measure and manage AI risk. You do not need to implement them in full to start; an inventory and a one-page policy already cover the first steps of the “Govern” and “Map” functions.
The short version
Shadow AI tells you where your people see value. Find it without blame, sort it by the data involved, give people approved tools for the use cases that matter, and check again every quarter.
If you want help with the inventory, the policy or an approved internal assistant, our AI governance and security practice does exactly this, and our free AI Opportunity Audit includes a first look at how AI is already used across your teams.
Sources
- NIST, AI Risk Management Framework 1.0 (January 2023) and the Generative AI Profile, NIST AI 600-1 (July 2024).