Shadow AI at Work: Your Staff Are Already Using It

Someone in your company pasted a customer contract into a chatbot last week to get a summary. Someone else uploaded a spreadsheet of sales figures to build a chart faster. A third person asked an AI tool to rewrite a difficult email to a supplier, including the parts about pricing.

None of them asked permission. None of them thought they were doing anything wrong, because they were not trying to break a rule, they were trying to finish a task. This is shadow AI: generative AI tools used for work without IT knowing they exist.

Why shadow AI is different from shadow IT

Unauthorised software is not new. People have been installing their own file sharing tools and note apps for two decades. What changed is the shape of the risk.

With an unapproved app, the problem is usually where the data sits. With shadow AI, the problem is what people type in. An employee does not upload a database, they paste the three paragraphs that matter, and those three paragraphs are often the most sensitive part of the document.

The barrier is also lower than ever. There is nothing to install and nothing to approve. A browser tab and a free account is the whole setup, which is exactly why banning it does not work.

What actually goes wrong

Be careful here, because the popular version of this story is exaggerated. Reputable providers do not hand your text to strangers, and most business tiers exclude your data from model training entirely.

The realistic risks are more mundane and more likely:

• Consumer accounts on personal email. Nobody can revoke access when that person leaves, and the history goes with them.
• Free tiers with different terms. The consumer version of a tool often has weaker data protection than the paid business version of the same product.
• Confidentiality obligations broken quietly. If your contract with a client says their data stays within your systems, pasting it elsewhere is a breach whether or not anything leaks.
• Personal data in the wrong place. Employee records or customer details processed by a tool nobody assessed is a data protection problem.
• Answers taken at face value. A confidently wrong summary of a contract clause can be more expensive than any data leakage.

Banning it makes shadow AI worse

The instinct is to write a policy that forbids everything. It fails for a simple reason: the tools genuinely help, and people will not give up something that saves them an hour a day.

What a ban really achieves is silence. Usage moves to personal phones, where you have no visibility at all, and nobody tells you when something goes wrong because admitting it means admitting they broke the rule. You have not removed shadow AI, you have removed your own view of it.

The organisations handling this well do the opposite. They make the sanctioned route easier than the unsanctioned one.

A workable approach

Ask before you assume. Find out what people already use and why. The answers tell you which problems your current tooling fails to solve, which is useful information regardless.

Provide a proper tool. Give staff a business-tier assistant with a contract behind it. If your company runs on Microsoft 365, the built-in option keeps data within your existing agreements. Once a good tool exists, most shadow AI disappears on its own.

Write a short AI policy. One page, not fifteen. Name the approved tools, list what must never be pasted anywhere, and be explicit that a human checks anything that goes to a customer. A policy people can remember beats a document that only exists for auditors.

Draw the line around categories, not tools. Personal data, customer material covered by confidentiality, credentials, unpublished financials, anything under NDA. Clear categories survive the arrival of the next tool.

Invest in employee training, briefly. Twenty minutes with real examples from your own work beats a compliance module. Show a good use and a bad use side by side.

Make it safe to speak up

One cultural detail matters more than any control. People need to be able to say “I pasted something I probably should not have” without fear.

Early knowledge lets you notify a client, rotate a credential or correct a document while it is still cheap. Silence costs you that option. If your only response to a mistake is blame, you have chosen not to find out about the next one.

Let us give your team a safe alternative

Shadow AI is not a discipline problem. It is a signal that your staff want to work faster than their current tools allow, and that is worth taking seriously rather than punishing.

At EvolvingDesk we help you turn that into something manageable. We look at what is actually being used, put a properly licensed assistant in place inside your existing environment, write the one-page policy with you, and run the short session that makes it stick. Where it matters, we set the technical controls so sensitive data cannot casually leave.

Your team keeps the speed. You keep the oversight. Nobody has to pretend they are not using AI.

Want to know what your staff are already using? Plan a call and we will help you find out without turning it into an investigation.

Read more

• The IT foundation you need before adopting AI
• What is Microsoft Copilot and how can your business use it
• How to keep remote workers productive and secure

Further reading: the NCSC guidelines on secure AI use and the ENISA material on AI and cybersecurity.