Clinical AI needs
evidence in the workflow.

“AI in healthcare” is not one job.

A model that flags an area on a scan, a system that predicts which patients may need follow-up, and a tool that drafts a visit note affect different decisions. They need different measures of success and different checks for harm.

Start by naming the exact task, who uses the output, and what happens next. A score for image classification cannot tell you whether a triage workflow shortens delays, whether a draft note is accurate, or whether a screening pathway improves outcomes.

Accuracy on a dataset is not the same as benefit in care.

Retrospective studies can reveal whether a model recognizes patterns in previously collected records. Their usefulness depends on how data were selected, how labels were assigned, and whether patients or sites in the test set resemble the setting where the tool may be used.

Clinical evaluation asks a wider question. Does the system fit the workflow? Do the people using it see the right information at the right time? What errors follow, and how are they caught? The answer depends on the intended use, the comparison pathway, and outcomes that matter for that task.

MASAI tested a workflow, not a stand-alone model.

The Swedish MASAI trial randomly assigned 105,934 women to AI-supported mammography screening or standard double reading. In the AI pathway, the system helped triage examinations to one or two radiologist readings and highlighted suspicious findings for review. A radiologist remained part of every screen reading.

In the 2025 screening-performance analysis, cancer detection was 6.4 per 1,000 participants with AI support and 5.0 per 1,000 with standard reading. Screen-reading workload was 44.2% lower. Recall and false-positive rates were not significantly higher in the AI group.

These are useful operational and screening results from one defined programme, model, and comparator. They do not show that every medical AI tool works, or that screening results alone establish a change in long-term health.

The two-year result calls for precise wording.

The later MASAI analysis followed the primary endpoint: cancers diagnosed between screening rounds. Interval cancer rates were 1.55 per 1,000 participants in the AI-supported group and 1.76 per 1,000 in the standard-reading group. The result met the study’s prespecified non-inferiority margin; it did not establish that AI was superior on this endpoint.

The same paper reported higher sensitivity in the AI group (80.5% versus 73.8%) and the same specificity (98.5% in both groups). These measures describe performance in this trial. They do not by themselves establish a reduction in mortality or guarantee the same result in another screening programme.

The 2025 performance paper and this follow-up report on the same randomized trial, not two independent replications. More findings at the first screen, fewer readers, interval diagnoses, false positives, and long-term health are related questions, but they are not interchangeable outcomes.

Documentation assistants need their own trial.

A pragmatic randomized trial at one academic institution assigned 238 outpatient physicians across 14 specialties to one of two ambient AI scribes or usual care. Over the study period, the systems were used in 33.5% and 29.5% of visits in their respective groups. The primary measure was time spent writing notes, not diagnostic accuracy or patient outcomes.

One scribe was associated with a 9.5% decrease in time-in-note versus control; the other showed no significant change in that primary outcome. Clinicians reported occasional clinically significant inaccuracies with both systems. Possible improvements in survey measures of workload and burnout were secondary findings that the authors said need confirmation in larger, multicenter trials.

The result is neither a verdict on every scribe nor evidence that the tools change care quality. It is a concrete example of why the product, setting, actual usage, primary measure, and errors all belong in the same account.

Match the proof to the claim.

1. Intended use: What exact decision or task is the system meant to support?
2. People and setting: Who was studied, at which sites, and how closely does that match the intended population?
3. Comparator: Was the AI pathway compared with the care people would otherwise receive?
4. Outcome: Does the primary measure fit the claim—accuracy, time, missed cases, follow-up, or a patient outcome?
5. Trade-offs: What happened to false positives, missed cases, workload, and the need for additional review?
6. Use in practice: How often was the tool actually used, and what happens when people override or cannot use it?
7. After release: How are errors, population shifts, software changes, and site-level performance monitored?

Regulatory status answers a different question from clinical effectiveness. For example, the FDA’s public list covers AI-enabled devices authorized for marketing in the United States, evaluates a device for its intended use, and explicitly says the list is not comprehensive.

Evidence belongs to the system around the model.

For the wider evaluation toolkit, read how to read AI benchmarks ↗ and how to measure AI bias ↗. For AI in another scientific workflow, explore AI for scientific discovery ↗.