Somewhere tonight a registered nurse will finish a twelve-hour shift, get home, open a chatbot on a personal laptop, and ask it to turn eleven years of med-surg experience into a resume. Nobody at her hospital approved the tool. Nobody trained her on it. Nobody will ever know she used it.
Hundreds of thousands of American nurses have taught themselves to work this way over the past year. The behavior is being catalogued under a security heading.
The security framing answers a real question. It is simply not the only question the behavior raises.
The Numbers on Unsanctioned Use
The scale is no longer in dispute, though the exact figure moves with the methodology and the population.
The 2026 PagerDuty Shadow AI Survey, conducted by Wakefield Research among 1,250 office professionals in non-IT roles at companies with a minimum annual revenue of $500 million across the United States, United Kingdom, Australia, and Japan, found that 66% had used AI tools at work despite believing this was not permitted under company policy. Okta’s AI Agents at Work 2026 report, surveying knowledge workers across seven countries, puts unapproved use at 52%. A BlackFog study of employees in the US and UK landed at 49%, and found something more interesting about who is comfortable with it: 69% of presidents and C-suite members were fine with the practice, prioritizing speed over privacy.
The response has been uniform across the vendor ecosystem. Inventory the tools. Deploy an approved alternative. Monitor the endpoints. Train on policy. It is a containment posture, and it treats the behavior as leakage to be plugged.
Containment addresses the data risk. It also discards the thing that is arguably more valuable than the data is risky, which is that a meaningful fraction of the workforce has been running unfunded, self-directed skill acquisition on the most consequential technology of the decade, and almost none of that learning is being captured.
Two Different Behaviors Wearing the Same Label
Before going further it is worth separating what these surveys are actually measuring, because shadow AI covers two populations that call for different responses.
The first is knowing defiance. A rule existed, the worker believed it applied, and the worker proceeded anyway. That is the population containment programs are designed for, though its motive is rarely what containment assumes.
Okta’s own data on motive is the tell. Among workers using unapproved tools, 80% said it was simply easier to use their own accounts, and 78% said their team already did it and it was considered normal. The approval process was too slow or difficult for 57%. Almost none said they were unaware that approval was required. That is not a compliance failure. It is a product failure, in which the sanctioned path lost to the unsanctioned one on the merits.
The second population is quieter and probably larger still: workers adopting tools inside a vacuum, where no rule was ever articulated clearly enough to defy. Healthcare is the cleanest place to observe it, because the employer baseline is close to zero and the confound disappears.
Incredible Health’s 2026 State of Nursing Report surveyed 2,240 U.S. registered nurses alongside platform data covering roughly 1.5 million healthcare professionals. Nurse AI use went from 15% to 44% in twelve months.
Now the employer side of the same population. Only 8% of nurses say their employer has communicated a clear AI strategy for how AI should be applied to their work. One in five say AI tools simply appear with no plan or explanation. And 46% received no AI training in the past year.
Some of that adoption did arrive through procurement. Epic AI, used by 28% of nurses who use AI, is the hospital’s own system, and ambient documentation comes through a contract. But the tool mix shows where the weight sits: 65% of nurses using AI reach for general-purpose assistants, with Microsoft Copilot at 40%, ChatGPT at 37%, and Gemini at 24%. Only 34% use anything healthcare-specific at all.
A workforce is teaching itself on consumer tools while sitting inside organizations with clinical informatics departments, enterprise software agreements, and formal training infrastructure. The training infrastructure was pointed elsewhere.
What the Self-Taught Cohort Learned
Here is where the reframe earns itself, because the same dataset shows the learning was real and directionally correct.
On whether AI will help the workforce, regular users say yes at 67% against 35% of non-users. On whether it will hurt, regular users say yes at 16% against 38% of non-users. The gap is wide, and it does not run the direction most governance programs assume.
The conventional enterprise sequence is to reassure the workforce first and expect adoption to follow. This population ran it backwards, and adoption produced the reassurance. The report frames it as a flywheel: comfort grows with use, and use grows with comfort. Workers on the sidelines are the most anxious ones.
They also learned the limits, which is the part the security framing has no way to register. Among nurses using AI, 83% say the output is rarely or only sometimes accurate enough to act on without checking it first. Half say their most recent use saved little or no time. They found the reliability ceiling on their own, without a policy document pointing them at it.
And the training effect is measurable on top of that. Among nurses without proper training, 16% save more than an hour a day. Among nurses with proper training, 24% do. Half again as many people clearing the same bar, from training alone, on identical tools.
The Cheapest Intervention Nobody Runs
There is one lever in the healthcare data with a measurable return, and it costs almost nothing.
Nurses who were consulted when AI tools were selected trust AI at 74%. Nurses who were never consulted trust it at 38%. On whether they will actually use a tool that has been deployed, the consulted group runs 81% against 62% for those who were never asked. Same tools, same institutions, same clinical work. The variable is whether anyone asked them first.
Fewer than one in ten nurses say frontline staff are always included in tool selection.
That is a doubling of trust available to any organization willing to put users in the room during selection, and almost nobody is doing it. It is the cheapest input in the entire governance stack, it requires no procurement and no budget line, and it is skipped almost universally by organizations that then express surprise at the adoption pattern that follows.
The Governance Implication
If bottom-up is the default shape of enterprise AI adoption rather than an aberration, then most AI governance programs are aimed at a sequence that is not happening.
Governance built for top-down rollout assumes a procurement event, a defined user population, a training rollout, and a monitoring period. Governance for bottom-up adoption needs different primitives: a mechanism for surfacing what people are already using, a sanctioned path that beats the unsanctioned one on the merits rather than on policy, users involved early enough that trust exists before deployment, and training that starts from the assumption that the user already has habits.
That last one is where most programs break. Policy training tells a workforce what it may not do. The people who taught themselves do not need to be told what a chatbot is. They need to be taught the thing they cannot learn alone, which is how to catch a confident wrong answer. Nearly half of nurses have received nothing on that subject.
The healthcare case is stark only because the employer baseline is so low. Every sector has some version of it. Every organization contains a group of people who are better at this than the people writing the policy, and none of them will volunteer it in a survey that frames the question as a violation.
The 66% who used AI believing it was prohibited did not do so to expose data. They did it to finish the work. That is a talent signal being processed as an incident report.
