Banning ChatGPT did not stop your staff using it. It stopped you seeing it.

Somebody in your organisation pasted a client contract into ChatGPT this week. Probably several people did. They were not being reckless. They had a deadline, the tool was open in another tab, and it did in thirty seconds what would have taken them an hour.

If you banned ChatGPT last year, that is still what happened. The difference is that it happened on a personal account, possibly on a personal phone, and you have no record of it.

A ban does not remove the risk. It removes your view of it.

How many of your staff are using AI you did not approve?

Probably about a third, and the ones using tools you have banned are telling you something. Deloitte's survey of 25,000 UK workers, carried out by Ipsos between May and June 2026, found 63% use generative AI for work and 31% use it without their employer knowing. Nearly one in ten admitted using a tool their employer had banned or would disapprove of.

Many of them are paying for it themselves. 17% of the people using AI at work pay for at least one tool out of their own pocket, which Deloitte estimates adds up to £958 million a year. People do not spend their own money on something unless it saves them real time. The same survey put the average saving at 70 minutes a week.

31% of UK workers use generative AI without their employer's knowledge, and around half of those using it have had no formal training on how to use it safely. — Deloitte, September 2026

That last figure is the one to dwell on. Half of the people using these tools have been given no guidance at all, so they are making their own decisions about what is safe to paste.

Where the data actually goes

The best evidence on this comes from browser telemetry rather than surveys, because people under-report what they paste. LayerX's Enterprise AI and SaaS Data Security Report 2025, drawn from monitoring real enterprise browsers, found that 77% of employees who use AI tools paste data into them, and 22% of those pastes contain personal or payment card data.

The number that matters for a ban is this one: 82% of those pastes came from unmanaged personal accounts. Not the corporate tenant with data retention switched off. A personal login, governed by consumer terms, that your security team cannot see, audit or delete.

LayerX sells browser security, so read the framing with that in mind. The direction is consistent with every other source, though, and it matches what you would expect. Block the tool on the corporate network and people do not stop. They switch to their phone, or log in with their own email.

What it costs when it goes wrong

IBM's Cost of a Data Breach Report 2025 was the first to measure this separately. One in five organisations in the study had a breach involving shadow AI, and those breaches cost an average of $670,000 more than the rest. 65% of them exposed customer personal data, against 53% across all breaches.

The governance figures are worse. 63% of the breached organisations had no AI governance policy or were still writing one. Of those that did have a policy, only 34% checked whether anybody was following it.

So most organisations either have no rule, or have a rule and no idea whether it works. A ban is often the second kind.

Why bans fail

A ban is a policy that asks people to choose between their deadline and a rule they do not understand. The deadline wins. It wins more reliably when 64% of weekly users are already worried their manager thinks AI could replace them, which is another Deloitte finding. People who feel that way do not volunteer that they are using AI. They hide it.

The pattern we see most often in professional and financial services goes like this. The firm bans public AI tools on the grounds of client confidentiality, which is a reasonable concern. It does not offer an alternative, because choosing one feels like a big decision and nobody owns it. Six months later, junior staff are drafting with AI on their phones and retyping the output, senior staff assume the ban is working, and the confidentiality risk the ban was meant to prevent is now larger and invisible.

The uncomfortable part is that the ban mostly protects the people who wrote it. It gives leadership something to point at. It does very little for the client whose contract was pasted into a personal account at eleven at night.

What works instead

None of this means anything goes. It means the control has to be one people will actually use.

• Give people a sanctioned tool. An enterprise licence with training on your data switched off, single sign-on and retention you control. If the approved option is as easy as the unapproved one, most people will take it. If it is worse, they will not.
• Write down what can go in it. Not a 30-page policy. A page that says which data is fine, which needs anonymising first and which never goes in any AI tool. Make it specific to your work: client names, health data, unreleased financials.
• Amnesty before enforcement. Ask people what they are already using and for what, without consequences. That list is the most useful AI strategy document you will ever get, because it tells you where the time savings are. Most of it will be drafting, summarising and searching, and much of it may not need AI at all.
• Check the supplier's terms, not the marketing. Where is the data processed, is it used for training, how long is it kept and can you delete it? These are the same questions we suggest you ask any AI vendor before signing.
• Then monitor. Once there is a good sanctioned route, blocking the unsanctioned ones is fair and enforceable. Doing it the other way round is how you end up with a ban nobody follows.

The real decision

The question in front of you is not "should we allow AI". Your staff answered that without you. The question is which tool you sanction, for which work, with which data, and that is a much smaller and more answerable decision than it feels.

It is also the kind of question our sanity check is for: £1,250, one day, and a written answer on which tool fits the work your people are already doing, what data should stay out of it, and what you do not need to buy at all. You can see what the write-up looks like on what you get.