Do you know which AI tools your team is using?

Ask most business owners whether their staff use AI at work and you get one of two answers. Either “no, we haven’t rolled anything out yet” or “a bit, I think, for drafting emails”.

Both are almost always wrong.

Microsoft’s UK research found that 71% of employees have used unapproved consumer AI tools at work, with just over half doing so every week. Set against that, only around a third of organisations have any formal policy governing AI use at all. The CIPD’s Labour Market Outlook has tracked the same gap from the HR side, with the large majority of UK organisations reporting employees using AI tools while guidance lags well behind.

The shorthand for this is shadow AI, and it is worth being clear about what kind of problem it actually is. It is not a technology problem. Nobody has been hacked. It is a data problem and a contract problem, and those are the two that end up in front of a regulator or a client.

What it actually looks like

The abstract version of this risk never lands. The specific version does.

In social care, a support worker at the end of a long shift pastes a set of rough case notes into a free chatbot and asks it to tidy them up. Those notes contain a service user’s name, their address, their medication and a description of an incident. That is special category personal data, processed on a consumer platform, under terms nobody in the organisation has read, with no record that it happened.

In a legal firm, a paralegal drops three pages of a client’s witness statement into a summarising tool to save an hour. The retainer says client information will be kept confidential. The firm’s professional indemnity position assumes the same. Neither anticipated an American consumer service.

In recruitment, candidate CVs go into a tool to be reformatted or scored. Those CVs belong to individuals who gave their data for one purpose. If a tool is ranking or filtering candidates, you have also wandered into automated decision making, which the ICO takes a particular interest in.

In every case the member of staff was trying to do their job faster. None of them thought they were doing anything wrong, because nobody had told them otherwise.

Why a ban does not work

The instinctive response is to block everything. It is also the response that guarantees the problem gets worse.

Bans do not remove the tools, they remove your visibility of them. Staff move to personal phones and personal accounts, which is precisely where you have no logging, no data protection agreement and no ability to answer a client asking what happened to their file. Research on shadow AI consistently finds the same two reasons people use unapproved tools: the approved option does not exist, or the approved option is worse. A ban addresses neither.

There is also a competitive point. The time savings are real. If your people have found a way to draft a proposal in twenty minutes instead of two hours, the answer is not to take it away. It is to give them a version of it you can stand behind.

Visibility first

You cannot write a sensible policy about something you cannot see.

Start with what your existing systems already tell you. Microsoft 365 tenants have reporting that shows which applications staff are signing into with work accounts. Firewall and DNS logs show what is being reached from the office. Expense claims show what people have quietly started paying for themselves.

Then ask. Not as an audit with consequences attached, but as a genuine question, because people who fear being disciplined will simply stop telling you. Most businesses that run this exercise find between five and fifteen tools in use that leadership had never heard of, and a couple of them turn out to be genuinely useful.

Then give them something approved

Once you know what people are doing, the fix is usually to provide a sanctioned equivalent rather than to write rules about the unsanctioned one.

If you are already on Microsoft 365, the Copilot route is the shortest path, because it sits inside the tenancy you already have. Microsoft’s data protection commitments state that prompts, responses and data accessed through Microsoft Graph are not used to train the underlying foundation models, and that the service operates under the same contractual terms as the rest of Microsoft 365.

Two caveats worth knowing before anyone gets comfortable. First, the consumer version of Copilot is a different product under different terms, and staff signed in with personal accounts are not covered by any of the above. Second, Copilot respects existing permissions, which sounds reassuring until you remember how many SharePoint folders have been quietly overshared for years. It does not create new access, it just makes years of sloppy permissions instantly searchable. Tidying that up is part of the work, not an optional extra.

Other tools can be approved too. The test is whether there is a business agreement in place, whether you know where the data goes and whether you could explain the arrangement to a client without wincing.

Then one page of policy

A twenty page AI policy will not be read. One page will.

It needs to cover four things. Which tools are approved and for what. What must never be pasted into any AI tool, named specifically for your business, so case notes, client files or candidate CVs rather than “confidential information”. That AI output is a draft and a named human remains responsible for what goes out. And who to ask when someone wants to use something new, with a clear commitment that asking will not get them into trouble.

The ICO’s guidance on AI and data protection sets out how UK GDPR obligations apply, and the NCSC’s overview of AI and cyber security covers the security side. Neither requires you to become an expert. Both are worth a read before you write your page.

Your clients are about to ask

This is the part that turns a governance question into a commercial one.

Security questionnaires have already started carrying AI questions. Which tools do you use, what data goes into them, who is accountable for the output, do you have a policy. We wrote about the broader trend in what to do when customers ask about your IT security, and AI is now the fastest growing section of those documents.

“We don’t really use AI” is not a good answer, because it is usually not true and the other side often knows it. A short, honest policy and a list of approved tools is a much better one, and it takes an afternoon to produce.

Confidentiality heavy environments can be run well. Our work with David Auld and Co Solicitors is a good example of how the underlying controls hold up when the information involved is genuinely sensitive.

Where to start

Visibility, approved tooling, one page. In that order, because a policy written before you know what people are actually using is a document about an imaginary business.

Our business IT solutions team can run the discovery exercise, work out whether Copilot or something else is the right sanctioned option for how you actually work, and help you get the permissions tidied up first.

Get in touch and we will start with finding out what is already going on.