Shadow AI: Your Staff Are Already Pasting Company Data Into ChatGPT. Now What?
Shadow AI starts when staff adopt AI before the company does. How to find the tools in use, what data went in, and how to move people to approved tools.
In this article
Shadow AI is any AI tool your employees use for work that the company has not approved, set up or can see. Often it is a personal ChatGPT account with a customer email, a spreadsheet or a contract pasted into it. If that is happening at your business, a ban is the wrong first move. Treat shadow AI as a policy gap: find out which tools are in use, decide what data never goes in, then move people onto approved, company-owned AI accounts and train them to use those well. A business with 5 to 100 employees can do most of that in a month.
This is general information, not legal advice. If you think customer financial, health or ID information has already gone into a personal AI account, bring in your attorney early.
Shadow AI is a policy gap, not a discipline problem#
Your staff are not pasting data into chatbots to cause trouble. They are trying to get work done faster, and nobody gave them an approved way to do it.
In Microsoft and LinkedIn's 2024 Work Trend Index (opens in a new tab), a survey of 31,000 knowledge workers in 31 markets, 78% of AI users said they bring their own AI tools to work, rising to 80% at small and medium-sized companies. And 52% of people who use AI at work are reluctant to admit using it for their most important tasks.
Most are guessing at the rules. In the U.S. results of a University of Melbourne and KPMG study (opens in a new tab) (1,019 U.S. respondents, surveyed November 2024 to January 2025), half of U.S. workers said they use AI tools at work without knowing whether it is allowed, and 46% said they had uploaded sensitive company information and intellectual property to public AI platforms.
When half your team does not know the rules, the rules were never written down. Punishing people for guessing wrong only teaches them to hide it, and hidden use is the kind you cannot fix.
What makes shadow AI risky?#
The bigger risk is the account, not the AI. A personal account runs on consumer terms your company never agreed to. You have no admin view, you cannot pull the history, and everything in it leaves when the employee does.
Consumer and business plans treat your data differently. Anthropic lets people on its Free, Pro and Max plans choose (opens in a new tab) whether their chats are used to improve Claude, keeps that data for five years when they allow it, and says those terms do not apply to commercial services such as Claude for Work. On the business side, Microsoft states that prompts, responses and data accessed through Microsoft Graph aren't used to train (opens in a new tab) the foundation models behind Copilot.
Breach research shows what missing rules cost. IBM's Cost of a Data Breach Report 2026 (opens in a new tab), a Ponemon Institute study of 602 breached organizations, found that the share of security incidents involving shadow AI more than doubled, to 43% from 20% the year before, and 49% of those incidents led to data loss or compromise. Most of the breached organizations (68%) lacked AI governance to manage AI or detect shadow AI. IBM's sample is breaches of 2,590 to 115,380 records, so read the percentages as a direction, not a forecast for your business. The lesson still carries: the gap was missing rules.
Shadow AI is also not just chatbots. It includes browser extensions that read the pages you open, meeting note-takers that join calls, and apps people sign into with their work Microsoft or Google account. Microsoft's documentation notes that, by default, users can consent to apps (opens in a new tab) for permissions that don't need an administrator, such as letting an app access their mailbox.
How to find shadow AI in your business#
Start with what your systems already record, then ask people. The goal is an inventory, not a list of names.
1. Check which apps are connected to work accounts#
In Microsoft 365, open Enterprise apps in the Microsoft Entra admin center (opens in a new tab) and filter to Enterprise Applications to see the non-Microsoft apps in your tenant. The user consent settings decide whether staff can connect new ones on their own. In Google Workspace, go to Security, then Access and data control, then API controls, and open Manage Third-Party App Access (opens in a new tab). It shows which third-party apps have reached Google data, how many users each has, and which services it requested.
Look for AI writing assistants, meeting note-takers and "chat with your files" tools.
2. Pull a list of browser extensions#
Google's Chrome Enterprise Core (opens in a new tab) manages Chrome at no additional cost, and its Apps and extensions usage report (opens in a new tab) lists installed extensions, how many browsers have each, and their permissions. On Microsoft Edge, if the extension blocklist policy (opens in a new tab) is not set, users can install any extension, so set one.
3. Read the card statements#
Look for recurring charges from AI vendors on company cards and in expense reports. A personal subscription the company reimburses is company work in an account the company does not control.
4. Run a no-blame survey#
The most complete list comes from asking. Say up front that nothing reported counts against anyone, then ask:
- Which AI tools have you used for work in the last 90 days?
- What do you use each one for?
- Were you signed in with a personal account or a work account?
- What kinds of information went in: customer, employee, financial, health, contracts or passwords?
- Which tool would you hate to lose?
The last answer matters most. The work people already hand to AI is usually your best list of places where AI saves time.
The voluntary NIST AI Risk Management Framework (opens in a new tab) asks for the same thing: mechanisms to inventory AI systems (opens in a new tab) (GOVERN 1.6). For a small business, the log review and survey above are that inventory.
What data should never go into an AI tool?#
Some information stays out of every AI tool unless you have approved that specific tool and use in writing:
| Never paste | Why |
|---|---|
| Social Security, driver's license and state ID numbers | Named in Wisconsin and Florida breach laws |
| Bank account and card numbers, and their access codes | Named in both breach laws |
| Passwords, MFA codes, or a username or email with its password | Florida's law covers login credentials |
| Health, medical and health insurance information | Florida's law covers it, and HIPAA limits who covered health providers and plans can share it with |
| Employee pay, reviews, discipline and medical leave | Confidential personnel information |
| Client files covered by a contract, NDA or professional duty | Your contract sets the rules, not the AI vendor's terms |
Wisconsin's breach law, Wis. Stat. § 134.98 (opens in a new tab), covers a name linked to an unencrypted Social Security, driver's license or state ID number, financial account number or access code, DNA profile or biometric data. When a business knows someone it did not authorize acquired that information, it must notify the people affected within a reasonable time, not to exceed 45 days. Florida's § 501.171 (opens in a new tab) defines a breach as "unauthorized access of data in electronic form containing personal information" and requires notice no later than 30 days after a breach is determined or there is reason to believe one occurred. Neither law mentions AI. If you are a HIPAA covered entity, 45 CFR 164.502(e) (opens in a new tab) lets you share protected health information with a business associate only when you obtain satisfactory assurance that it will safeguard the information, documented in a written contract (usually called a business associate agreement). A personal chatbot account gives you neither.
If sensitive data already went in#
- Write down what went in, which tool, which account and roughly when.
- If it includes breach-law personal information, health information or client-confidential material, call your attorney before deleting anything.
- Then delete those chats and turn off any setting that lets the vendor train on them.
- Move that work into a company account going forward.
Deleting a chat clears it from the person's history. What the vendor keeps after that depends on its terms.
How do you move people to approved AI tools?#
Give them a company account as good as the one they use now#
Blocking without a replacement fails. If you run Microsoft 365, web-based Copilot Chat is included with eligible subscriptions at no extra cost (opens in a new tab), and Microsoft says it includes enterprise data protection (opens in a new tab) for eligible work and school accounts. If your team prefers ChatGPT or Claude, buy business seats. Our ChatGPT, Claude and Grok buyer's guide compares seat prices and data settings.
Write a one-page shadow AI policy#
Our free AI acceptable use policy template covers approved tools, a red, yellow and green data list, human review of AI output, and same-day reporting of mistakes. Set an amnesty date when you publish it: nothing before that date counts against anyone.
Train people on the approved tool#
A policy says what not to do. Training shows people what to do instead, using their own work. Hands-on AI training for business teams, built around real tasks from your business, gives people a reason to reach for the approved tool first.
Then tighten the controls#
Once people have a better option, close the side doors: allow user consent only for apps from verified publishers (Microsoft's recommendation), block unapproved AI extensions, and review connected apps every quarter. That work belongs with your other cybersecurity services, next to security awareness training. If HIPAA, PCI DSS or the FTC Safeguards Rule applies to you, your AI rules belong inside that program too, and our compliance management work includes policy development for those frameworks.
A 30-day shadow AI plan#
| Week | What to do |
|---|---|
| 1 | Pull the connected-app and extension lists. Send the no-blame survey. |
| 2 | Choose approved tools and buy seats. Take any sensitive exposures to counsel. |
| 3 | Publish the one-page policy. Hold a 30-minute team meeting. Set up accounts. |
| 4 | Train on the approved tool. Restrict app consent and extensions. Book a quarterly re-check. |
Frequently asked questions#
What is the difference between shadow AI and shadow IT?#
Shadow IT is any software or device used for work without approval, such as a personal file-sharing account. Shadow AI is the AI slice of it. It carries more risk because AI tools only work when you feed them information, so normal use means pasting company data in.
Should we just ban ChatGPT at work?#
A ban with no alternative tends to push the use onto personal phones, where you cannot see it. Approve a business-grade tool, write down what never goes in, and block specific tools once people have something better.
Can I see what my employees typed into ChatGPT?#
Not in a personal account. The company has no admin access to it, which is the core problem with shadow AI. In a company-owned plan you own the accounts, and review options depend on the product. For Microsoft Copilot, Microsoft documents that admins can view stored prompts and responses (opens in a new tab) with Content search or Microsoft Purview.
Is pasting customer data into ChatGPT a data breach?#
It can be. Wisconsin's and Florida's laws turn on unauthorized acquisition of, or access to, specific personal information, and neither mentions AI. Whether a paste into a personal account qualifies depends on what went in and the vendor's terms. Write down the facts and call your attorney quickly, because the notice clocks run up to 45 days in Wisconsin and 30 days in Florida.
Get an AI readiness review#
If you want an outside review of your AI readiness before you choose tools, our AI adoption consulting starts with an assessment of how work actually gets done in your business. From there it covers tool recommendations, a written acceptable use policy that says what staff may put into which tools, and a rollout with staff training. Our Eau Claire office serves businesses across the Chippewa Valley, and our Holmen and Fort Myers offices serve the Coulee Region and Lee County. Start with a free discovery call.
