Someone on your leadership team mentions, almost in passing, that half the sales team pastes customer emails into ChatGPT to draft replies. Nobody approved it. Nobody told you. It has probably been going on for months.
That is shadow AI in the workplace: people using AI tools for company work without the company knowing, approving, or setting any rules. It is not a sign that your team is reckless. It is a sign that capable people found a tool that saves them time and nobody gave them a better path.
Most advice says to write a policy. That is right, and we cover it in writing an AI use policy. But a policy written without knowing what people are actually doing will describe a company that does not exist. This piece is about the step that comes first, or alongside: finding out what is in use, stopping the few things that cannot wait, and bringing the rest into the open.
Why Shadow AI in the Workplace Happens
People are not hiding AI out of malice. They are hiding it, when they hide it at all, for ordinary reasons.
- The tool is free and takes two minutes to start using.
- The work is tedious, and the tool makes it faster.
- Nobody said they could not, and nobody said they could.
- They fear that admitting it will look like cutting corners, or like proof a machine can do their job.
That last reason matters most. Our piece on what employees fear about AI goes deeper, but the point for now is simple: if your first move is to hunt for offenders, the people who are using AI well will go quieter, and you will lose the one thing you need, which is honest information.
Adoption is real and broad. Census Bureau survey data from May 2026 shows 32% of US firms with 100 to 249 employees reporting AI use in business operations, and 37% for firms with 250 or more. That is the formal, company-reported number. The informal number, the one that counts individual employees quietly using personal accounts, is almost certainly higher. We do not have a verified figure for it, and you should be skeptical of anyone who quotes you one.
Find Out What Is in Use Without a Witch Hunt
You have two broad ways to learn what is happening. Use both, in this order.
Ask first
Start with a conversation, not a scan. Tell your leadership team and department heads that you are not looking to punish anyone. You are trying to learn what tools people find useful so the company can support them properly. Then ask three plain questions:
- What AI tools have you used for work in the last 90 days?
- What did you use them for?
- What kinds of information did you put into them?
Do not collect names in a spreadsheet to review later. Collect tools, tasks, and data types. The goal is a map, not a list of suspects.
You will hear things you did not expect. Some of it is harmless. Some of it is exactly what you feared. All of it is useful to know.
Then check what the systems already show you
Once people have had a fair chance to tell you, look at what your existing tools can tell you. This is not surveillance of individuals. It is checking your own work:
- Expense reports and company cards for AI subscriptions people are paying for themselves and expensing.
- Browser extensions and software installed on company devices, if your IT team already manages them.
- Single sign-on or email-login records, which can show which new services staff signed up for with work addresses.
- The settings in the AI features already built into the software you pay for. Many email, document, and meeting tools added AI features that may be switched on by default.
Whoever runs IT should do this, and should report patterns (which tools, how many users), not individual behavior, unless something serious turns up.
What to Stop Today
You do not need a finished policy to draw a few lines. Some uses should stop the day you learn about them, and you can say so without waiting for anything to be written.
Customer and client data
Names, emails, account details, order histories, support conversations. If it identifies a customer, it does not go into a consumer AI tool on a personal account. The reason is not that the tool is evil. It is that you have no agreement controlling where that data goes, who can see it, or whether it is kept. Our piece on AI security for business covers the risk in more detail.
Contracts and legal documents
Draft agreements, signed contracts, settlement terms, anything your attorney would call privileged. Pasting these into an open tool can undo confidentiality you are obligated to keep.
Anything regulated
Health information, financial records, employee records, anything covered by a regulation or a customer contract that restricts how data is handled. If you are not sure whether something is regulated, treat it as regulated until someone who knows tells you otherwise.
Credentials
Passwords, keys, and system access details never belong in an AI tool.
Say this list out loud to the whole company, in plain words, in the same week you start asking questions. Four categories fit on a card. You are not banning AI. You are marking the places where the cost of a mistake is too high to leave to individual judgment.
What to Allow in the Open
Here is where many owners overcorrect. Having found shadow AI, they ban everything, and the use goes further underground.
Instead, decide what you can allow now, provisionally, in the open. For most companies, the low-risk list looks like this:
- Drafting and editing text that contains no customer, contract, or regulated information.
- Summarizing public information or the person's own notes.
- Brainstorming, outlining, and first drafts of internal material.
- Help with formulas, spreadsheets, and routine analysis on non-sensitive data.
Attach two conditions. First, a person is responsible for whatever goes out under the company's name, so a person reads and checks the output. Our article on the verification layer for AI reliability explains why this matters. Second, anyone using a tool for a new kind of task tells their manager, so the map stays current.
Allowing the safe things openly does more for you than banning them. It gives people a legal place to stand, and it moves them from hiding to reporting.
How the Amnesty Conversation Goes
This is the hard part, and most owners skip it. If you have learned that people have been pasting sensitive material into AI tools, you need to talk to them. How you handle that conversation sets the tone for everything after.
Set the terms before anyone speaks. Say this: anything done before today, in good faith, without a rule against it, will not be held against anyone. From today, the stop list applies.
Own your part. You did not give them a policy, tools, or guidance. They filled the gap. Say that plainly and the room relaxes.
Ask, then listen. What problem were you solving? What would have made it easier to do this the right way? You will learn where your processes are slow and where the real demand for AI lives. Some of those answers will shape your first approved tools.
Be specific about what changes. Name the stop list. Name what is allowed. Name who to ask when something is unclear, and make that person approachable.
Handle the exceptions honestly. Amnesty covers honest mistakes made before a rule existed. It does not cover someone who keeps doing it after the line is drawn, or who hid a serious breach. You can say so without making the whole meeting a warning. If a customer's information was exposed in a way that triggers a legal or contractual duty to notify, that goes to your attorney the same day, separate from the amnesty talk.
There is a stewardship question underneath all of this. The people in your company trusted you with their livelihoods, and your customers trusted you with their information. Treating a mistake with grace while still protecting what you were trusted to hold is not a conflict. It is the job. We explore that more in our stewardship and AI framework.
Where This Leads
Think of discovery, the stop list, the allowed list, and the amnesty conversation as the groundwork under your policy. When you write it, you are no longer guessing. You have a real inventory, a few lines already drawn and explained, and a team that has seen you handle the first awkward conversation with fairness.
After the policy is published, the work changes again: training, questions, enforcement, and the first slips. We cover that in the first 30 days after you publish the policy.
Do This Today
Before you do anything else, take 20 minutes and write the four-item stop list on one page: customer data, contracts, regulated information, credentials. Send it to your leadership team with one line: "We are not banning AI. These four things stay out of it until we have rules. Tell me what you are using so we can do this properly."
That one message starts discovery, draws the lines that cannot wait, and signals that this is a conversation, not an investigation.
Your Next Step: The AI Use Policy Builder
When you are ready to turn what you learned into a written policy, use the AI Use Policy Builder. Bring the map you built: the tools in use, the stop list, and the allowed list.
If you would rather talk through your situation first, the Discovery Call is a 30-minute conversation at no cost and no obligation.