Your Team Is Already Using AI You Can’t See

How many of your team use AI at work that you haven’t approved?

If you’re guessing a handful, Microsoft has some news for you. When Censuswide asked 2,003 UK employees last October, 71% said they’d used unapproved consumer AI tools at work. Just over half do it every week.

Then in September, Deloitte found something even more telling. UK workers spend an estimated £958m a year of their own money on AI tools for their jobs.

That’s shadow AI. The NCSC defines it as “the use of AI technology which isn’t captured in an organisation’s approved systems and processes”.

Most businesses file it under security, and it does belong there. It’s also the clearest signal you’ll get of what your people need to do their jobs.

How Many People Are Doing It?

Every survey asks the question slightly differently, so the numbers jump around. Some count all staff, some only people who already use AI.

Deloitte’s is the lowest because it only counts people who already use AI. It still means 31% of them are doing it where their employer can’t see.

It’s Your Senior People

So who are they? You might picture a junior marketer sneaking ChatGPT into a campaign. The data points further up.

BlackFog’s January survey of 2,000 UK and US staff, reported by CIO.com, found 69% of presidents and C-suite executives put speed ahead of privacy. And CIPD’s 2026 Good Work Index found professionals (51%) and managers and directors (45%) are the heaviest users of generative AI of any group in UK work.

So if you’re about to sign a policy that bans unapproved AI, there’s a fair chance it would catch you.

Why People Hide It

If it’s that common, why does nobody mention it in meetings?

Slack asked 17,372 desk workers in 2024. 48% said they’d be uncomfortable telling their manager they’d used AI for common tasks. The top reasons were that it felt like cheating (47%), and fear of looking less competent (46%) or lazy (46%). The least common reason was that company policy discouraged it.

Deloitte’s UK survey found the same thing two years on. 23% of workers think there’s a stigma attached to using AI at work, and 64% of weekly users worry their manager will decide AI can do their job.

So the fix sits with managers. When a manager says “show me how you did that”, people open up. Slack found people comfortable telling their manager were 67% more likely to have used AI for work at all.

What Goes Into It

So what are people pasting in? Harmonic Security tracked 22.5 million prompts and uploads at US and UK organisations through 2025 and found 579,113 cases of sensitive data.

It’s also very concentrated. Six apps carried 92.6% of the sensitive exposure, and ChatGPT alone carried 71.2%. LayerX found 83.5% of users stick to a single AI tool.

And the paste habit is everywhere. LayerX’s browser data found 77% of AI users paste data into chatbots, about 22% of those pastes contain personal or payment details, and 82% come from personal accounts.

Why do personal accounts matter so much? OpenAI says that on its individual plans “we may use your content to train our models” unless the user switches it off, while its business plans don’t by default. The NCSC warns that data sent to consumer AI services may be “stored, retained or used to improve the service”. Your data ends up in a retention regime you didn’t choose, under an account IT can’t see or close.

It’s already reached the regulators. In May this year, a Pennsylvania bank filed what American Banker reported as the only SEC filing on record to use the words “unauthorized artificial intelligence”. An employee had put customer names, Social Security numbers and dates of birth through an unapproved AI app.

What It Costs

IBM’s 2025 Cost of a Data Breach report studied 600 breached organisations. One in five had a breach linked to shadow AI. Where shadow AI was high, it added as much as $670,000 to the average breach, on top of a global average of $4.44 million.

Customer personal data was exposed in 65% of shadow AI breaches, against 53% of breaches overall. And 63% of the breached organisations had no AI governance policy at all.

Gartner expects more than 40% of enterprises to have a security or compliance incident linked to unauthorised shadow AI by 2030.

The Skills You Never Build

Leaks get the headlines. The slower cost is what nobody in the business learns from hidden use.

If the use is hidden, the mistakes still get made and whatever the person learned stays with them. KPMG’s numbers show what that looks like across the UK.

Of UK workers who use AI at work, 58% have relied on its output without checking it, 54% have made mistakes in their work because of it, and 45% have used it to do a task so they didn’t have to learn how.

CIPD’s 2026 index found 66% of people who use generative AI at least monthly have had no AI training at all. Deloitte found 65% of UK users see no convincing lead from the top on how to use it. People have bought the tools themselves and are teaching themselves, quietly.

When the Shadow Tool Can Act

The new wave of personal AI agents, Meta’s Muse, OpenAI’s dots and Anthropic’s Claude Duet, can read your email, check your calendar and act on your behalf. Most aren’t officially available in the UK yet.

A chatbot leaks what you paste into it. An agent connected to a work inbox can reach everything in that inbox, all the time.

And it gets there with one click. The agent asks for access to Google Workspace or Microsoft 365, and the employee presses Allow. If your settings let staff approve third-party apps themselves, that’s the whole process. The access runs under the employee’s own login and can outlive their time with you. Microsoft only tightened its default in mid-2025, and that change covered apps asking for files and sites, on tenants that never changed the setting. By Microsoft’s own documentation a user can still, by default, let an app into their mailbox.

BlackFog’s survey, reported by CIO.com, found 51% of employees have connected AI tools to work systems or apps without IT knowing.

Security researcher Simon Willison calls the danger the lethal trifecta: an agent with access to private data, exposure to content from strangers and the ability to send things out can be turned against you by anyone who can get text in front of it. An agent on your inbox has all three by design.

It’s happened. In 2025 researchers showed the EchoLeak flaw could make Microsoft 365 Copilot leak internal data from a single email with hidden instructions, with no click from the victim. Microsoft fixed it. In January, researchers showed a calendar invite could get Google Gemini to copy someone’s private meetings into an event the attacker could read.

The NCSC’s warning is plain. If an attacker exploits a vulnerable agent, they get “the same data, services, and privileges” the agent holds.

Why Banning It Doesn’t Work

So should you just ban it?

Software AG’s 2024 survey of 6,000 knowledge workers in the US, UK and Germany found 46% would keep using personal AI tools even if they were banned outright. TrustedTech found 22% of UK employees would carry on even with disciplinary action on the table.

What does work is giving people a proper option. Netskope measures its own customers, and its 2026 report found personal-account use among generative AI users had fallen to 47%. Its retail customers show the swap most clearly: staff on ungoverned AI fell from 70% to 44% in a year, while use of approved tools rose from 40% to 73%.

The NCSC says shadow AI “is unlikely to disappear completely”, and that policies which can’t meet business needs push people towards unapproved tools.

If you have staff or customers in the EU, there’s one more reason. Article 4 of the EU AI Act, as amended in July, asks employers to “take measures to support the development of AI literacy” of their staff. Training is the plainest way to meet it.

What to Do on Monday

Shadow AI tells you your people want this badly enough to pay for it themselves. Give them the tool and the training before the next person buys their own.

So who on your team is paying for their own AI right now, and what would it take for them to tell you?

Sources & References

Frequently Asked Questions

Roy Murphy is co-founder of G3NR8. This post draws on 2024 to 2026 research on shadow AI, workplace AI use, data security and AI agents. Published 24 October 2026.