Shadow AI: it's rarely malicious, and that's exactly why it's hard to spot
Most of the company data leaving UK businesses through AI tools isn't being taken by someone with bad intentions. It's being pasted in by a good employee, on a deadline, trying to finish something faster.
Ask most business owners whether their staff are putting company data into ChatGPT, Claude, Gemini, Copilot or a dozen free AI writing tools, and you'll get one of two answers. Either "no, we don't allow that" or "probably a bit, but nothing sensitive."
Both answers have the same problem. They're not based on data. They're based on the absence of data — nobody has reported anything, so it must be fine.
This is what people mean by shadow AI: the use of AI tools inside a business without the knowledge, approval or oversight of whoever is responsible for its data. It's the direct descendant of shadow IT, but it moves faster, costs nothing to adopt, needs no installation, and leaves almost no trace on a traditional network.
The scale is bigger than most owners assume
Microsoft and LinkedIn surveyed 31,000 knowledge workers across 31 markets, including the UK, for their 2024 Work Trend Index. Two findings stand out.
First, 75% of knowledge workers were already using generative AI at work. Second — and this is the one that matters for anyone responsible for data — 78% of those users were bringing their own AI tools to work rather than using something their employer had provided. At small and medium-sized companies that figure rose to 80%.[1]
Put those together and the picture is uncomfortable but clear: in a typical business, the majority of AI use is happening on tools the business didn't choose, doesn't pay for, and can't see.
Why good people do this
It's tempting to frame shadow AI as a discipline problem. In our experience it almost never is. The same Work Trend Index research found 68% of people say they struggle with the pace and volume of work and 46% report feeling burned out.[1] An AI tool that turns a two-hour job into fifteen minutes isn't a temptation to those people. It's a lifeline.
Think about how the decision actually looks from the employee's desk:
A bid manager pastes last year's tender response into a chatbot to reword it for a new client. It contains commercial terms and named contacts.
Someone in HR uploads a spreadsheet of applicants to get a summary of the shortlist. That's other people's personal data.
A finance assistant drops a management accounts extract in to ask why a variance looks odd.
An operations lead installs a free meeting transcription tool that quietly joins every call in the diary, including the ones about redundancies.
Not one of those people is doing anything they'd describe as risky. Each one is being resourceful. The tool is free, it's in a browser tab, there's no download and no approval step, and the output is genuinely good. There is no moment in that workflow where anything says stop.
And there's a second, quieter reason this stays hidden. Microsoft found that 52% of people who use AI at work are reluctant to admit using it for their most important tasks, and 53% worry that using it makes them look replaceable.[1]
That's worth sitting with. Your best-performing staff have a direct personal incentive not to tell you which parts of their job AI is doing. So even a well-meant "just come and ask us first" policy runs into a wall. Asking permission means admitting use, and admitting use feels like admitting you're easier to replace.
This is why survey-based approaches — asking your team what they use — reliably underestimate the truth. You're not measuring behaviour. You're measuring what people are comfortable telling you.
What actually goes wrong
The risk isn't hypothetical, and it isn't only about the headline case of a model being trained on your data.
The data leaves your control, permanently
Once a document is pasted into a consumer AI service, it sits on someone else's infrastructure under their terms, not yours. You cannot reliably retrieve it, you cannot prove it was deleted, and you often cannot say who at that provider could see it. That's the concern Samsung cited in 2023 when it restricted generative AI use across the company after staff uploaded confidential internal source code — the company noted the data was stored on external servers, making it difficult to retrieve or delete.[3]
If Samsung, with its resources, found this out the hard way, it's a fair bet most SMEs haven't yet found out at all.
It costs real money when it contributes to a breach
IBM's 2025 Cost of a Data Breach report — conducted by the Ponemon Institute across 600 breached organisations — measured this directly. Organisations with high levels of shadow AI paid an average of £498,000 more per breach than those with low or no shadow AI. IBM described it as one of the largest single cost drivers in the 2025 study.[2]
Most UK organisations have no controls in place
The same IBM research found that only 31% of UK organisations had governance policies in place to manage AI use and prevent shadow AI, and 63% had no AI access controls at all. Globally, 97% of organisations that suffered an AI-related security incident lacked proper AI access controls.[2]
The UK Government's own data points the same way. The Cyber Security Breaches Survey 2025/2026 found that among UK businesses using AI, adopting it or actively considering it, only 24% had any cyber security practices or processes in place to manage the risks from it. The same survey found 43% of UK businesses reported a cyber security breach or attack in the previous twelve months.[4] Shadow AI doesn't cause most of those. But it widens the blast radius when one happens.
The compliance position is uncomfortable
Under UK GDPR, you remain the data controller for personal data your staff handle — including personal data they paste into a tool you've never heard of. The ICO has been explicit that adopting AI does not reduce your existing data protection obligations, and generative AI remains a stated priority area in its current strategy.[5] If you can't say what personal data has been shared, with which processors, under what lawful basis, that's a gap that exists whether or not anything has gone wrong yet.
If you trade into the EU, there's a second layer worth knowing about. The EU AI Act reaches businesses based outside the EU where the output of an AI system is used in the Union — so it can apply to a UK firm with EU customers, EU staff or EU operations, even with no EU entity. Its AI literacy duty under Article 4 applies to deployers of AI systems, not just the companies building them, and has been in force since February 2025. The Act's high-risk obligations were pushed back to December 2027 by the Digital Omnibus in July 2026, so there is more runway than there was — but a delay isn't a repeal, and the literacy duty and the Article 50 transparency rules are live now.[6]
We're not going to pretend the AI Act is an emergency for a UK business with no EU exposure, because for many of you it isn't. The point is narrower than that: every one of those obligations — literacy training, transparency, any deployer-side risk assessment, and increasingly the AI questions appearing in customer due-diligence questionnaires — starts by assuming you can list the AI systems in use across your business. That list is the thing most organisations don't have.
Why banning it doesn't work
The instinct is to write a policy prohibiting AI tools and circulate it. It's cheap, it's quick, and it feels like action. It also tends to fail, for three reasons.
It drives use onto personal devices. Block the tools on the corporate network and the work simply happens on a phone or home laptop, where you have less visibility, not more. The data still leaves — you just stopped being able to see it go.
It punishes your most productive people. The staff using AI are frequently the ones under the most pressure and delivering the most. A blanket ban asks them to do the same volume of work more slowly, and offers nothing in return.
It's unenforceable without visibility. A policy you can't measure compliance against is a statement of hope. And the numbers above suggest people won't self-report — 52% are reluctant to admit AI use on their most important work.[1]
What good looks like instead
The organisations handling this well aren't the ones with the strictest rules. They're the ones who can answer four questions on any given day:
Which AI tools are being used across our business right now? Not which are approved — which are actually in use, including ones nobody has told you about.
What kind of data is going into them? There's a real difference between someone rewording a blog post and someone uploading a customer database.
Where does that leave us legally? Which of those tools are processing personal data, and do we have any basis for that?
What's the sanctioned alternative? If you take a tool away, what do you give people instead so the work still gets done?
Notice that three of the four are visibility questions. You can't write a sensible AI policy until you know what's actually happening — otherwise you're regulating a business you've imagined rather than the one you have.
It's also worth saying that IBM's data cuts both ways on AI. UK organisations using AI and automation extensively in their security operations saw breach costs of £3.11m against £3.78m for those that didn't — and identified and contained breaches 42 days faster.[2] The technology isn't the enemy here. Ungoverned use of it is.
How Enhance IT approaches this
We treat shadow AI as a visibility problem before it's a policy problem. In practice that means four things.
The point of all of this isn't to catch anyone out. It's to make sure that the person who pastes a client contract into a free chatbot on a Friday afternoon is stopped by a system, not by their own knowledge of a policy they skimmed nine months ago.