Where healthcare AI has proof, and where it does not

Step 1 of 8

Seven places hospitals spend on AI

Documentation, administrative work, care transitions, diagnostics, the workforce, chronic disease and behavioral health.

Step 2 of 8

Sort them by how good the proof is

Left to right: early concept, single-site results, consistent pilots, peer-reviewed multi-site data. Only published results move a category along this line.

Step 3 of 8

Then by market momentum

Bottom to top: contested, steady, building, accelerating. Momentum combines investment, adoption and regulation, and a category's momentum and evidence often differ: some move fast on thin proof, and some well-proven ones move slowly.

Step 4 of 8

Proven and moving

Documentation and diagnostics have strong evidence and strong momentum. Documentation is the easier of the two to buy, though one vendor quotes bigger time savings than the first major study across several vendors found. For diagnostics, a 2026 study in routine practice found two radiologists more sensitive than two commercial AI systems at finding lung nodules.

Step 5 of 8

Momentum ahead of evidence

Chronic disease management and workforce tools are growing faster than the evidence behind them. Workforce relief cannot yet be bought as a single product, and chronic disease monitoring needs staff assigned to act on the readings. A narrow, measured pilot is the sensible step.

Step 6 of 8

Strong evidence, slow market

Care transitions has the strongest combined cost and outcome evidence on the map, and it is the only category here with a published result in which a racial gap in readmissions closed. Adoption is still slow. The programs that work pair an AI model that flags high-risk patients with staff who follow up with those patients, and those staff are the hard part to fund.

For hospitals in CMS's mandatory TEAM model, care transitions now affect payment directly, because TEAM counts spending through 30 days after discharge against each surgery's target. We publish a free position report for any participating hospital.

Step 7 of 8

How hard each one is to deploy

Bubble size is deployment readiness this year. Documentation is the easiest to deploy. Behavioral health is small for a reason: two states ban AI therapy chatbots outright, and in the first three months of 2026, 36 states introduced more than 70 bills to regulate AI chatbots.

Step 8 of 8

The verdicts

Two categories are rated Adopt, three Pilot, one Watch and one Hold. Select a bubble in the explorer below for the sources behind each rating.

Step 1 of 6

Six use cases

Discharge and transitions, dementia care coordination, ambient documentation, remote monitoring, prior authorization and emergency risk stratification. Each is scored on three outcomes that have to hold at the same time.

Step 2 of 6

Score the patient outcomes

This column asks whether patients are better off, so it is measured on patients. Clinician burnout counts in the next column instead.

Step 3 of 6

Score the operational relief

This column counts capacity freed, staff time returned and delays removed. The saved time has to show up somewhere you can measure it.

Step 4 of 6

Score the money

The money is counted after the cost of running the tool, including the staff who act on what it produces. Leaving that cost out makes this column look better than it is.

Step 5 of 6

Two of six meet the bar on all three

Discharge and transition optimization and dementia care coordination reach Moderate or better on all three. That matters because a project that is strong on two outcomes and weak on the third can launch and still not be renewed.

Step 6 of 6

Ambient documentation does not meet it yet

It scores Emerging on patient outcomes and on money. The first major study across several health systems and vendors found it saved about 16 minutes of documentation time per 8 scheduled patient hours, which the study's senior author called modest. If you pilot it, measure your own numbers before it starts, so you have something to compare against afterward.