Eighty-three percent of clinicians are using artificial intelligence at work with no guidance from their employer. No formal policy, no recommended tool, no approved workflow. That figure comes from a survey of 1,823 clinicians across 25 countries conducted in May 2026 by Heidi, published this month as its Pressure Points report, and it describes a profession that has adopted a technology faster than the institutions employing it could write the rules.
Five in six. In a field where the introduction of a new blood pressure cuff triggers a validation protocol, that number is remarkable.
This is not experimentation anymore
The instinct is to read unsanctioned use as clinicians tinkering at the margins. The data says otherwise. Eighty-six percent report using AI daily or several times a week, and 57 percent now describe it as a consistent part of how they work. That is not curiosity. That is dependence, formed without oversight.
What they are using it for is unglamorous and entirely predictable. Eighty-eight percent of respondents named documentation as their most time-consuming administrative burden, and that answer held across every geography, specialty, and career stage in the survey. Paperwork is the one thing every clinician in the world agrees is eating their day, and a tool that drafts a note in thirty seconds instead of six minutes does not need an implementation committee to justify itself.
They know exactly what the risks are
The easy assumption about shadow adoption is that people using tools without permission have not thought about the downside. The survey rules that out. Sixty-eight percent named hallucinations and accuracy as their leading concern, ahead of patient privacy at 59 percent, over-reliance at 47 percent, and erosion of clinical judgement at 41 percent.
Clinicians have ranked the failure modes correctly and in a sensible order. They are using the tools anyway, because the administrative pressure is immediate and concrete while the risk is probabilistic. Anyone who has faced a queue of unfinished notes at the end of a shift understands that trade, whether or not they would defend it in a governance meeting.
The governance gap is documented
The same week the Heidi figures circulated, a Nitro report found that only 36 percent of healthcare organisations have an actively enforced policy for AI document security. Set those numbers next to one another and the picture is stark. Roughly five in six clinicians are working without institutional guidance, and roughly two in three institutions have no enforced policy covering the documents those clinicians are feeding into these systems.
That is not primarily a clinician problem. It is a governance failure, and the correction is not a ban. Prohibition against a tool that is already daily practice for 86 percent of your workforce produces concealment rather than compliance, and concealment is the condition under which a hallucinated dosage actually reaches a chart.
Patients are further along than the debate assumes
The consumer side of this has shifted quietly. A Sogolytics survey found that just 49 percent of consumers now prefer a human interaction for healthcare and medical consultations. That is a slim minority favouring a human being, in medicine, which would have been an implausible result a few years ago.
Usage patterns point at why. A Druid report found that 29 percent of healthcare AI conversations happen outside business hours, and that 87 percent resolve without escalation to a person. Availability appears to be doing much of the work here. A system that answers a question at eleven at night is competing against nothing, not against a clinician, and most of those exchanges never needed a clinician in the first place.
What the numbers ask of hospital leadership
The practical reading of the Pressure Points data is that the adoption argument has been settled by the workforce and the governance argument has not been started by the employers. Clinicians have identified documentation as the pain point, selected tools themselves, and built those tools into daily routine. The institutions have contributed policy vacuum.
The gap worth closing is narrower than it looks. Approving specific tools, defining what patient data may enter them, and requiring clinician review before AI-drafted content reaches a record would address the accuracy concern that 68 percent of respondents already flagged themselves. None of that requires a new technology. It requires deciding.
Watch whether the 83 percent figure moves in next year’s survey. If it drops, it means health systems finally wrote the policies. If it holds, it means a profession spent two full years running critical infrastructure on tools nobody officially approved.
For more coverage of artificial intelligence in healthcare, visit Mylistingo.
