Most employees use AI tools nobody approved. The UK's cyber agency says the fix is a better approved one.

I've run enablement for new AI workflows: training, office hours, and materials for non-technical teams. One lesson keeps coming back. Watch what tool people reach for when nobody's checking. That tells you more than the policy does.

Microsoft UK put a number on it last year. In a Censuswide survey of 2,003 UK employees, 71% had used AI tools their employer never approved, and 51% did it every week. The reasons were mundane. 41% said it's what they use in their personal life. 28% said their company doesn't offer a work-approved option at all.

In September, the UK's National Cyber Security Centre published its take. It didn't call for a tighter blocklist. It said shadow AI is "unlikely to disappear completely," so the goal is to reduce the risk rather than assume it can be eliminated. It also said shadow AI tends to show up when security policies don't meet business needs. Its advice was to build a security culture where people can talk openly, and to provide secure alternatives.

The risk is real, which is why this matters. PagerDuty commissioned Wakefield Research to survey 1,250 office professionals at large companies in four countries. Two-thirds had used AI tools at work they believed their company didn't allow. 34% had put customer data into public AI tools, and 31% had shared financial information or confidential documents. And 75% said they'd likely look for a new job that offered better AI skills development.

Put those together and the pattern is clear. People aren't going around policy because they enjoy the risk. They're doing it because the approved option doesn't do the job, or doesn't exist. Running QA on LLM training data taught me to read behavior at volume as data. Every time someone opens a public model instead of the internal tool, that's a labeled example of what the internal tool is missing. Most companies file it under compliance risk. I'd file it under product research.

Treat it like research and the response changes. First, find out what's actually in use and for what. Use a survey people will answer honestly, since an IT incident report only shows who got caught. Second, talk to the five heaviest users of unapproved tools and ask what job the outside tool does that the approved one doesn't. Third, hand that answer to whoever owns the approved tool, and log it as a product requirement.

That said, this doesn't mean anything goes. Some data should never touch a consumer model, and the NCSC lists breaches, lost intellectual property and regulatory failures as the cost of company or customer data landing in the wrong tool. Draw that line plainly and explain why. In my experience, people follow a rule they understand far more often than one that just says no.

It's the same thing I saw running enablement. A mandate without a workflow behind it gets worked around. Shadow AI is that lesson showing up as a security headline.

A blocklist tells people what they can't use. Ship the tool they went around you to find, and the policy starts enforcing itself.