How Do I Stop Employees Pasting Confidential Data Into AI Tools?
You don't stop it with a ban. You stop it by giving people something better than the toaster, then governing it. Seven moves that cut shadow AI without killing productivity.
Short answer: you don't stop it with a ban. You stop it by giving people an approved tool that's better than the one they'd sneak in, forcing company sign-in, keeping the crown jewels on private AI, and governing the lot. Bans just move the toaster to their phone.
The uncomfortable truth
Right now, someone in your company is pasting something they shouldn't into a chatbot. Not because they're malicious. Because they're busy, the tool is good, and nobody gave them a better one.
LayerX's 2025 enterprise report found 77% of employees paste data into generative AI tools, and 82% of that pasting happens through unmanaged personal accounts. (TechRepublic) Your security team can't see personal accounts. Your DLP can't see personal accounts. Your board definitely can't see personal accounts.
That's not a leak. That's a sieve with a subscription.
Why bans don't work
Samsung tried it. After engineers pasted source code into ChatGPT in 2023, the company banned generative AI tools on company devices and said it would build its own. (Bloomberg) Notice the second half of that sentence. The ban was the pause. Building their own was the plan.
A ban without an alternative just teaches people to use their phone.
How much does it cost when it goes wrong?
IBM's 2025 Cost of a Data Breach report found breaches involving high levels of shadow AI cost an average of US$670,000 more, and 63% of breached organisations had no AI governance policy or were still developing one. Of the organisations that did have policies, only 34% regularly audited for unsanctioned AI. (Help Net Security)
So: most companies don't have a policy, and most of the ones that do don't check it. Toastie is thrilled.
The seven moves
- Give them a better toaster. An approved AI assistant that's genuinely good, connected to your documents, available in the tools people already use. If the approved tool is worse, you've lost before you start.
- Company sign-in, always. SSO for every AI tool used for work. Personal accounts on company work: blocked, politely and firmly.
- Sort the data into four buckets. Public, internal, confidential, crown jewels. Tell people, in one page, what can go where.
- Keep the crown jewels at home. Board papers, customer and patient data, source code, legal files: run those on private or in-country AI you control.
- Watch the clipboard, not the people. Browser and endpoint controls that warn before sensitive data is pasted into an unapproved tool. Coach first, block second.
- Train for judgement, not fear. Twenty minutes, real examples, the fridge test: would I be fine if this sat in a stranger's fridge for a year?
- Keep the evidence. An AI inventory, risk tiers, approvals and a quarterly check for unsanctioned tools. That's the folder you hand the auditor, the customer and the board.
What good looks like in 90 days
- Weeks 1 to 2: find out what's actually being used (it's always more than you think).
- Weeks 3 to 6: roll out the approved tool and company sign-in, publish the one-page policy.
- Weeks 7 to 10: move the most sensitive workflow onto private AI.
- Weeks 11 to 12: audit, report to the board, repeat quarterly.
The bottom line
You can't out-police human nature. You can out-design it.
Don't ban AI. Own it. Then make owning it the easiest option in the building. If you want a hand, the AI governance guide shows how I set this up, and the self-check tells you where you stand in two minutes.
Quick answers
What is shadow AI?
Should we ban ChatGPT at work?
How common is employees pasting data into AI tools?
Does shadow AI make data breaches worse?
Get the next one before your toaster does.
New issues of The AI Glittch land on LinkedIn first. Subscribe there, and follow me for the bits too short for a newsletter.

AI entrepreneur and venture builder. Builds private, governed AI for boards, executives and governments. Writes The AI Glittch. Has never trusted a toaster.