TL;DR: Somebody in your business has already pasted something into a free AI tool that you would rather they had not. A supplier contract, a customer list, a difficult email, a spreadsheet with names in it. They did it to get an hour back, not to be reckless, and they did not mention it because there was nothing to mention. Banning it does not work. Giving people one tool that is safe to use, and one page saying what never goes in, does.
This is not hypothetical
The biggest study anyone has done on this asked more than 48,000 people across 47 countries what they actually do with AI at work. Just under half admitted putting company information into free public AI tools. Fifty-seven per cent said they pass AI-generated work off as their own without mentioning it. Only four in ten said their employer had given them any guidance on the subject at all. That was KPMG and the University of Melbourne, published in 2025, and you can read the global report yourself.
The UK government's numbers point the same way. In the Cyber Security Breaches Survey 2025/2026, 31% of businesses were using AI, adopting it, or actively considering it. Of those, just 24% had any security process in place to manage the risks. Roughly three quarters of the businesses using AI have not thought about the security side at all.
None of that is surprising if you have ever watched somebody try to write a difficult email at five o'clock on a Friday.
What people are actually doing with it
Not "harnessing generative AI to transform workflows". The real list looks more like this:
- Pasting a supplier contract in and asking what the notice period is.
- Dropping in a customer spreadsheet and asking for it to be tidied up.
- Feeding in an angry email and asking for a polite reply.
- Uploading meeting notes and asking for the actions.
- Asking for a quote, using last year's quote as the example.
- Photographing an invoice and asking what it says.
Every one of those is a sensible thing to want. Most of them also involve somebody's personal data, somebody's commercial terms, or both, leaving your business through a browser tab on a personal account.
Why banning it does not work
Two reasons, and the second one is the important one.
The first is that you cannot enforce it. The tools are free, they work in a browser, and half your team have them on their phones already. Blocking a few domains on the office network does nothing about the phone in somebody's pocket or the laptop at home.
The second is that a ban makes people hide it. That is what the 57% figure above is really telling you. If using AI is against the rules, nobody asks whether this particular thing is safe to paste in, because asking means admitting. You end up with the same usage and none of the visibility. We have never once seen a ban reduce AI use. We have seen plenty of them reduce honesty about it.
The four things that genuinely go wrong
You cannot get it back. Once a document has been pasted into somebody else's system on somebody else's account, there is no recall button, no legal hold, no export. The realistic best case is asking the person to delete their chat history and taking their word for it.
You cannot answer the questions you would be asked. Under UK GDPR your business is still the controller for personal data your staff hand to a third party, whether or not you knew about it. If a customer asks what happened to their information, or the ICO asks, "one of our team pasted it into a chatbot on a personal account" does not end the conversation. The ICO's guidance for organisations is the plain English starting point if you want to read the official version.
Your client contracts may already forbid it. If you do work for larger companies, look at the confidentiality clauses in your last three contracts. A good number now say client data may not be put into third party AI services without written consent. The person who signed that clause is almost never the person doing the pasting.
The answer is sometimes confidently wrong. AI tools invent things. Case references, clauses, figures, dates. They do it in the same tone they use when they are right, which is what makes it dangerous. A wrong number in an internal note is a nuisance. A wrong number in a quote you sent a customer is a different conversation.
"Does it train on my data?" is not quite the right question
The question that matters is which account, on which plan, under whose terms.
The free consumer version, the paid consumer version and the business version of the same product are not the same product. Business and enterprise plans from the major providers generally commit not to train on your content and sit under a proper data processing agreement. Free consumer tiers generally do not offer you those commitments, and in any case your business has no agreement with anybody, because your employee signed up with a personal email address.
So the working assumption is simple. Anything typed into a personal account on a free tier is outside your control, because it is. If you want a better answer than that, get people onto an account you control, and read the terms of that specific plan rather than the marketing page.
What good actually looks like
Pick one tool and pay for it. If your business runs on Microsoft 365, Copilot is the obvious starting point, because it works inside your own tenant and respects the permissions you already have. If your team have settled on something else, buy the business tier of that instead. The licence cost is trivial next to the alternative, which is everybody quietly using their own.
Make the approved tool the easy one. People do not route round a policy out of spite. They route round it because the sanctioned option is slower or they have not been shown how to use it. Half an hour of proper training does more for compliance here than any amount of policy.
Write one page, not a binder. Name the three categories that never go into any AI tool: personal data about identifiable people, anything covered by a client confidentiality clause, and credentials or bank details. Then one rule about output: check it before it goes out with your name on it. That is most of the value.
Say what happens when somebody gets it wrong. If the answer is "you will be in trouble", you are back to hiding. The answer that works is "tell us the same day and we will sort it out", because a mistake you hear about on Monday is fixable and one you hear about at renewal is not.
Write down which tools are approved, and let people ask for more. A short authorised list plus a route to request an addition beats a blanket no, because it turns shadow use into a request you can assess.
If you want a starting point, we publish our own AI acceptable use policy as a free template. It is twenty-four sections, written in plain English, and it covers the awkward bits most templates skip: meeting recorders joining calls uninvited, automation tools quietly posting data to an API, voice cloning, and what happens to prompts when somebody leaves. Copy it, cut what does not apply, put your name on it.
Finding out what is already being used
Do not start with an audit. Start with a conversation, because the fastest way to find out what people are using is to ask them in a way that does not sound like a disciplinary. Say you are choosing a tool to pay for and you want to know what they already like. You will get an honest list in ten minutes.
Then the technical side, which is worth doing anyway:
- Check which third party apps have been granted access to your Microsoft 365 accounts. Staff often sign into an AI service with their work account and consent to permissions on the way through. That list lives in your admin centre and is usually more interesting than people expect.
- Look at browser extensions on managed devices. AI writing assistants and meeting recorders install as extensions and can read the pages they sit on.
- Look at expense claims and card statements. Small monthly subscriptions in the £15 to £25 range tend to be somebody solving a problem on their own.
- Decide about meeting recorders specifically. A bot that joins calls and transcribes them is recording your clients as well as your staff, and that has consent implications that go beyond IT.
None of this needs to be a project. It is an afternoon, and it turns a vague worry into a list.
Where to start
Ask the question this week. Not "is anyone using AI", which gets you a room of blank faces, but "what have you tried, and what did it actually help with". Whatever comes back, you are better informed than you were, and you will have three or four honest examples to build a one page policy around.
If you would like a hand deciding which tool to standardise on, getting Copilot switched on properly, or checking what your staff accounts have already been connected to, get in touch or ring us on 01584 517234. We do this for businesses across Shropshire, Herefordshire, Worcestershire and Powys, and it is a short job with a genuinely useful answer at the end of it.