43% of dentists, 81% of physicians, and nobody counted the policies

We build software. We are not lawyers and nothing here is legal or compliance advice. Everything below is sourced and linked at the foot of the page, and what any of it means for your practice is a question for your own counsel or compliance advisor.

What is actually published

The American Medical Association fielded its 2026 Physician AI Survey between 15 January and 2 February 2026. 1,692 physicians. 81% report awareness or use of AI in their practice, against 66% in 2024 and 38% in 2023. The average number of use cases per physician went from 1.1 in 2023 to 2.3 in 2026.

The American Dental Association's Health Policy Institute runs a quarterly panel of US dentists. As of mid-2026, 43.3% use AI for at least one task, and another 26.4% say they plan to. The largest single category is imaging and diagnostics at 22.8%. Then insurance verification at 13.6%, explaining findings to patients at 13.2%, front desk check-in and business analytics at 10.1% each.

The interesting part is not adoption

It is where the training isn't. The AMA asked physicians how much AI training they had received from any source. 27% have had none at all. Broken out by setting, the share who have had some training runs 76% in a hospital, 71% in group practice, and 55% in solo practice — the lowest of every setting listed. Of everyone who had received any training, only 11% described it as "a lot".

So the smaller the practice, the less likely it is that anyone has been taught anything. That is not a surprise. It is worth having in a number rather than as a feeling.

The one dimension that goes the other way

The AMA asked how physicians expect AI to affect ten different things: work efficiency, diagnostic ability, clinical outcomes, patient convenience, cognitive load, stress and burnout, value-based care, revenue, the patient–physician relationship, and patient privacy.

Nine of the ten skew positive. Work efficiency runs 78% helpful against 7% harmful. Patient privacy is the only one where more physicians expect harm than help — 41% harmful against 13% helpful. The report flags it as the single net-negative area, and it is the only one it flags.

Then there is the number that names the problem. Concern about patient privacy runs at 42% for AI tools the institution provided, and 71% for tools it did not. Same physicians, same worry, nearly double when nobody procured the thing.

The figure that does not exist

Neither report says how many practices have written down which AI tools staff may use.

We read all fourteen pages of the AMA report to be sure. Its methodology section lists "governance, liability, and regulatory trust factors" as one of the survey's domains, and the published findings from that domain are about what physicians want from regulators — clear liability frameworks, post-market surveillance, oversight. There is no figure for whether the practice itself has a policy. The ADA page reports usage and intent to adopt, not governance.

We are not going to fill that hole with somebody else's number. The equivalent evidence for law firms has been published and is widely quoted; it is not evidence about medical or dental practices and borrowing it would be the sort of thing this blog exists to catch other people doing.

So: we cannot tell you what share of small practices have an AI policy, and as far as we can find, neither can anyone else. What is published is that adoption is high, that training is thinnest at exactly the smallest practices, and that the tools clinicians worry about most are the ones that arrived without a purchase order. Those three sit together in a way that is difficult to read as anything other than a governance gap, but that is a reading, and we would rather label it as one than dress it up as a statistic.

One more thing about the sources

We pulled the AMA numbers out of the PDF directly, page by page. The first attempt was an automated summary of that same file, and it came back with 2,000 physicians, a May 2023 field date, 39% adoption and 3.5 use cases per physician. Every one of those is wrong. Nothing about the summary looked wrong — it was fluent, formatted, and confidently specific.

That is why the source list below links the report rather than a write-up of it, and it is the same failure mode this whole series is about: a fluent answer produced by a machine that nobody checked.

What to do

Ask one question at your next staff meeting: which AI tools does anyone here use for work? Write the answers down. That is the whole exercise, and it takes about ten minutes.

Not to police it. To find out. In most practices this will be the first time anybody has asked, and the answers are more interesting than people expect. It is a ten-minute item for whoever chairs the meeting, and what you want back is the list itself, not a summary of it. You cannot write a policy about a list you do not have, and you almost certainly do not have the list.

The cost of not asking is simply that you do not know. Everything else in this series runs off that list, and none of it can start until somebody has written it down.

Once the list exists, the next post in this series is about what to do with it. If you would rather have someone go through it with you, that is one of the ways we work. It is not the point of this post.