For most businesses, the honest answer is no. Not because nobody would care, but because nothing today tells you when it happens. A staff member pastes a contract, a client list or a pricing sheet into a public AI tool to save twenty minutes, gets a useful answer, and moves on. No log, no alert, no record it ever happened.
This is not a hypothetical. It is the default state of most businesses right now, and the data on how often it happens has become clearer over the past year.
What the current data shows
A run of recent surveys converge on the same picture from different angles. PagerDuty's 2026 survey of 1,250 office professionals found that two thirds had used AI tools at work even though they believed doing so was not permitted. Cyberhaven's 2026 report, drawn from observed data movements rather than a survey, puts a frequency on it: the average employee enters sensitive data into an AI tool roughly once every three days.
- 2 in 3 office professionals have used AI tools at work they believed were not permitted (PagerDuty, 2026).
- 60% of IT teams say they cannot see the specific prompts or requests employees make to generative AI tools (Cisco Cybersecurity Readiness Index, 2025).
- Every 3 days is roughly how often the average employee enters sensitive data into an AI tool (Cyberhaven, 2026).
Gartner's read on where this goes is blunt: by 2030, it predicts more than 40% of organisations will suffer a security or compliance incident caused by the use of unauthorised AI tools. This is not a future risk. It is a present one that most businesses cannot currently see.
Why it happens even in careful teams
Nobody sets out to leak a contract. The pattern is almost always the same: a real deadline, a genuinely useful tool sitting one tab away, and no sanctioned alternative that is as fast. Ban the tool and the behaviour does not stop, it just stops being visible. Say nothing and the exposure compounds quietly, one paste at a time, across every person on the team.
The businesses that close this gap do not do it by winning a policy argument. They do it by giving staff a governed way to get the same speed without the exposure, and by building enough visibility to know when that boundary is being tested.
Three controls that actually close the gap
- A governed assistant staff will actually use. Shadow AI thrives in the absence of a sanctioned alternative. A governed assistant that is as fast as the consumer tool, connected to the systems people already work in, removes the reason to reach for ChatGPT on a browser tab nobody is watching.
- Visibility into what is actually happening. You cannot govern what you cannot see. Basic visibility into which AI tools are in use and what kind of data is moving through them turns an invisible risk into a manageable one. On the Cisco figure above, it is the control most IT teams say they lack.
- A written answer to “does this train a public model”. Every credible AI vendor can answer this in writing. If a tool cannot confirm your data never trains a model outside your control, that is the question to press on before anything else, not after.
Where to start
Start by finding out what is actually happening today, not what policy says should be happening. Expect a gap between the two. From there, the fix is staged: a governed assistant first, for fast, visible wins, then the connected foundation underneath it.