Adoption has outrun preparation. The most recent physician survey data shows use rising sharply while training lags, and the fix is not more technology.
The American Medical Association surveyed 1,692 physicians between 15 January and 2 February 2026. 81% reported using AI in their practices, up from 38% in 2023. 92% wanted more education and training in AI, and 27% said they had received no training on AI use cases from any source. Separately, Bain found that fewer than 10% of MedTech professionals in Asia Pacific combine deep technical AI skills with a robust understanding of the healthcare ecosystem. Adoption is moving considerably faster than capability.
The AMA's Physician Survey on Augmented Intelligence, fielded between 15 January and 2 February 2026 across 1,692 physicians in a range of specialties, found 81% using AI in their practices. In 2023 the same measure stood at 38%. Use has more than doubled in three years.
That is a fast curve by any standard, and it happened without a corresponding investment in preparing the people on it.
In the same survey, 92% of physicians reported wanting more education and training in AI. 27% said they had received no training on AI use cases from any source at all.
Read those two figures together. This is not a workforce resisting AI. It is a workforce asking to be equipped and largely not being equipped. The appetite is not the constraint.
Physicians also signalled what would build their confidence. 55% wanted involvement in implementation decisions so they could evaluate the clinical evidence themselves, and most held that clear liability frameworks would strengthen trust in healthcare AI. Both are governance requests rather than technology requests.
Bain's November 2024 analysis of the region found that fewer than 10% of MedTech professionals in Asia Pacific possess both deep technical AI skills and a robust understanding of the healthcare ecosystem. Bain described this as a significant bottleneck to advancing the region's MedTech AI capabilities.
The dual-fluency point is the one worth sitting with. Plenty of people understand healthcare. Plenty understand AI. The scarcity is in the overlap, and the overlap is precisely where deployment decisions get made.
The instinctive response to a training gap is to buy training. We would push back on the version of that which most organisations reach for first.
A one-day awareness session moves nobody from 0 to competent. It produces the feeling of having addressed the problem, which is worse than leaving it visible. The 27% who have had no training will not be meaningfully served by a lunchtime webinar, and the 92% asking for more education are not asking for a slide deck about what a large language model is.
What works, in our experience, is sustained cohort learning where people build something themselves. The moment someone has assembled a working agent, seen it fail in an instructive way, and had to decide whether its output was trustworthy, they can evaluate a vendor's claims. Before that moment they are guessing, however many overviews they have attended.
The second thing we would change is who gets trained. Most programmes target clinicians. The AMA data suggests clinicians are already asking. The people who more often lack the fluency, and whose decisions are more consequential, are the executives approving procurement and the board members signing off on risk. A physician using an AI scribe badly affects one clinic. A leadership team buying the wrong platform affects the institution for five years.
Our practical advice: find out how many people in your organisation could interrogate a vendor's validation study and reach an independent conclusion. If the answer is a handful, that is your real exposure, not the training percentage.