Shadow AI: your employees already use ChatGPT at work. What to do about it

Your office already has AI in it

Walk past the order desk, the AP clerk or the inside sales rep and look at their browser tabs. There's a good chance one of them is ChatGPT, Claude, Gemini or Copilot. Maybe they're cleaning up an email to a customer. Maybe they pasted in a supplier's price list to compare it. Maybe they're asking how to write an Excel formula.

That's shadow AI: AI tools people use for work without the company choosing them, paying for them or setting rules for them. It isn't a fringe thing. It's normal.

• In Microsoft and LinkedIn's 2024 Work Trend Index, 78% of AI users said they bring their own AI tools to work. At small and medium-sized companies it was 80%. The survey covered 31,000 knowledge workers in 31 markets in early 2024, so it's office staff, not the shop floor, and it's now more than two years old.
• MIT's 2025 study of business AI, as reported by Fortune, found that only 40% of companies had bought an official AI subscription, while employees at over 90% of companies regularly used personal AI tools for work. That study was based on interviews with 52 organizations and a survey of 153 senior leaders, so read it as a strong signal, not a census.
• Gallup's tracking shows 15% of U.S. employees used AI daily in their role as of May 2026, and 30% used it a few times a week or more.

The same Microsoft survey found something more telling: 52% of people who use AI at work are reluctant to admit using it for their most important tasks. So the owner who says "nobody here uses that stuff" may simply not have been told.

Why people go around you

It's rarely rebellion. Usually nobody gave them a tool or a rule, and the free one was one click away.

Gallup asks U.S. employees whether their company has communicated a clear plan for using AI. The share saying yes has gone up, but slowly.

U.S. employees who say their company has shared a clear AI plan

Gallup Panel, percent saying yes, May 2024 to August 2026.

1. May 202415%
2. May 202522%
3. Aug. 202524%
4. Nov. 202526%
5. Feb. 202625%
6. May 202625%
7. Aug. 202626%

Source: Gallup, AI in the workplace indicator.

Three in four employees say they haven't heard a clear plan. When there's no plan, people make their own. That's how a customer's drawing ends up in a free chatbot account tied to someone's personal Gmail.

Where the real risk is

The risk isn't that AI is evil or that it will take over the order desk. For a mid-size company, it comes down to three plain problems.

1. Data goes somewhere you don't control. When someone pastes a customer PO, a price list or an employee's details into a personal account, that data now sits under that person's account terms, not your company's. Business plans are different. Anthropic, for example, says that by default it does not use inputs or outputs from its commercial products, such as Claude for Work, to train its models. Consumer plans have their own, separate terms. That difference is the main reason to pay for a business plan at all.

2. Breaches get worse. IBM's 2025 Cost of a Data Breach study, run by the Ponemon Institute, looked at 600 organizations that had a breach. One in five reported a breach due to shadow AI, and only 37% had policies to manage AI or detect shadow AI. Those with high levels of shadow AI saw breach costs about $670,000 higher on average than those with little or none. Shadow AI incidents also exposed more customer personal data (65% vs. 53% overall) and more intellectual property (40% vs. 33%). Keep in mind every company in that study had already been breached, and IBM says the AI-related group is a small part of its sample.

3. Nobody checks the work. A chatbot can draft a great reply to a customer. It can also get a lead time wrong with total confidence. When AI use is hidden, nobody reviews the output, and nobody knows which emails or quotes it touched.

The best-known warning is still Samsung. In 2023, after internal data was accidentally leaked to ChatGPT, Samsung temporarily restricted generative AI tools on company devices. A 100-person distributor isn't Samsung. But a customer's drawings or pricing matter just as much to that customer.

You can't ban your way out of a tool that's one click away. Give people a better one.

Why a ban rarely works

The gut reaction is to block ChatGPT on the company network and send a memo. It feels safe. It usually isn't.

• Phones get around it. A rep can type the same question into an app on their own phone, which is even harder to see.
• The need doesn't go away. People use these tools because they save time on real work. The MIT study, as Fortune reported it, found shadow AI often delivers better returns than formal AI projects, while official projects stall.
• It pushes use further out of sight. With 52% of AI users already reluctant to admit using it on key tasks, a ban mostly teaches people to hide it better.

To be fair, there are cases where a hard block is right. If you work under defense rules like ITAR or CMMC, handle medical data, or have customer contracts that bar outside tools, you may need strict limits on where data can go. Even then, the goal is a controlled tool, not no tool. A ban with no replacement just moves the problem.

What to do instead: a one-page plan

You don't need an AI committee. For a 40-400 person company, this fits on one page and can be done this month.

Ask first. Make it clear nobody is in trouble. You want to know what people are using and why. The answers tell you where the real demand is. In most offices it's writing emails, summarizing long threads, cleaning up spreadsheets and reading documents.

Pick one tool and pay for it. One approved tool on company accounts beats five personal ones. When someone leaves, you shut off their access. Their chats don't leave with them.

Write the data rule in plain words. For example:

• Green: public info, your own drafts, general questions, formulas.
• Yellow (approved tool only): customer emails, POs, price lists, internal reports.
• Red (never): passwords, bank details, Social Security numbers, health info, anything a customer contract says can't leave your systems.

Check what's connected. Shadow AI isn't only chat windows. It's also add-ons and apps that ask for access to read mail or files. Those are worth a look on their own. We cover that in questions to ask any AI vendor about your data.

Fix the task, not just the tool. If three people on the order desk paste customer POs into ChatGPT to pull out line items, that's a sign. The real fix is a process that reads the PO, drafts the order in your ERP and has a person approve it, with the data staying in systems you control. Our guide on ChatGPT vs. custom AI for business processes covers when a chat tool is enough and when it isn't.

Where the other side has a point

Some owners will say the risk is overblown, and they're not entirely wrong. IBM's own report says AI-related breaches are still a small slice of its sample. Most shadow AI use is low-stakes: rewording an email, fixing a formula, summarizing a meeting. And a business plan isn't magic. It protects how a vendor uses your data. It doesn't stop someone from pasting the wrong file, and it doesn't check the answer.

That's why the plan above isn't about fear. It's about knowing what's in use, putting the sensitive work in a tool you control, and keeping a person on anything that goes to a customer or into your ERP.

What this looks like when it's done

A month from now, you want to be able to answer four questions in one sentence each. Which AI tool do we use? What data can go in it? Who reviews what it produces before it goes out? And which repeated task did we turn into a proper process instead of a copy-paste habit?

If you can answer those, you don't have a shadow AI problem anymore. You have an AI tool your team chose, with rules they understand.