AI Tools for Physicians: What Is Actually In Daily Use

There are two separate worlds of medical AI and they barely overlap.

One is the FDA device list: hundreds of imaging tools, each cleared for a narrow indication, deployed by departments after a procurement process. The other is the software physicians actually touch every day, almost none of which is a regulated device.

The second world is larger, moves faster, and gets far less scrutiny. Understanding why is the most useful thing a clinician can know about this category.

The regulatory line runs through review

Software that drafts a note, summarizes a chart, answers a reference question, or replies to a patient message generally falls outside the medical device definition, because a clinician reviews the output before it affects care. The clinician is the safeguard, and the safeguard is what makes the exemption defensible.

That is why this category reached real adoption in about three years while the cleared device list grew one submission at a time. No premarket submission, no predicate, no clinical evidence requirement.

It also means nobody reviewed these tools before they reached your clinic. Evaluation is entirely the buyer's job, and the buyer is frequently a health system committee rather than the physician who has to live with the output.

Ambient documentation: the category that won

Ambient documentation listens to the encounter and drafts the note. It has the highest adoption of any clinical AI, and the reason is not technical.

Documentation burden is the one problem in medicine where physicians agreed on the diagnosis before anyone offered a treatment. It is measurable, universally resented, and directly implicated in burnout. A tool that removes some of it does not need to prove clinical benefit to get adopted.

It also has the lowest failure cost here. A badly drafted note gets caught during the review the clinician was already performing. What degrades it is background noise, multiple speakers, and language switching, and the time saving evaporates if the draft needs heavy rewriting rather than light editing.

The differentiators between vendors are mostly operational: how deep the EHR integration goes, whether the note lands in the right fields, and how much editing the draft actually needs. See the ambient documentation records.

Clinical reference: the category to watch carefully

This is where the regulatory exemption is doing the most work. A system answering "what is the first-line treatment for this" is influencing a clinical decision, and the argument that clinician review makes it a non-device rests entirely on that review actually happening.

The better tools use retrieval-augmented generation: find relevant sources first, then answer constrained to those sources, with citations attached. That makes the answer checkable, which is the minimum bar for anything touching a clinical decision.

Grounding reduces fabrication without eliminating it, and it introduces a second failure point. If retrieval misses the relevant paper, the model answers confidently from what it did retrieve, and the result looks exactly as well-cited as a complete answer. A citation proves a source exists. It does not prove the source supports the claim attached to it.

Ask any vendor here what the system does when it does not know, and whether it will say so. Browse the reference tools.

Inbox drafting: the newest burden

Patient messaging volume grew enormously after portals became standard, and no corresponding time was created to answer it. Inbox burden now rivals documentation as a burnout driver, which is why message drafting arrived immediately behind ambient documentation.

Published early evaluations have been mixed on time saved and more positive on perceived effort and message quality. That is a real finding rather than a disappointing one ... reducing the cognitive cost of composing a careful reply is worth something even when the clock does not move much.

There is an unsettled question here the field keeps not answering. Empathetic phrasing generated by a model and sent under a physician's name is a disclosure issue, and nobody has decided what patients should be told.

Coding and documentation integrity: the uncomfortable one

Tools that review the chart and surface undocumented diagnoses sit on a genuinely uncomfortable line. Surfacing a real but undocumented comorbidity improves the accuracy of the record. The same mechanism, tuned slightly differently, is a machine for generating billable diagnoses.

The tool is not the variable. The threshold is. Accepting a suggestion places a diagnosis in a permanent record with downstream consequences for the patient's insurability and future care, so suggestion acceptance rates are a metric worth watching for a reason.

What to do with all of this

Treat the absence of regulation as information rather than as reassurance. Ask the same questions you would ask of a device: what happens when it does not know, what does the vendor do with your data, and what is the comparator.

The full evaluation framework is in how to evaluate a clinical AI tool. The tool records themselves are in the physician AI tools directory.