Your note-writing AI should know more than what it hears

Ambient AI scribes promised to fix one of medicine's oldest problems: the note that eats away at the clinician's evening. Point a microphone at rounds, let the AI listen, and out comes a complete draft. No more typing at 9 p.m. No more pajama time.

It sounded too good to be true. Turns out, it mostly was.

A four-month pilot in the pediatric intensive care unit at a quaternary children's hospital put a commercially available ambient scribe to the test during multidisciplinary rounds. The results, published in an American Academy of Pediatrics blog earlier this year, are worth every clinician's attention — not as a cautionary tale about AI, but as a clear map of where the technology needs to get better.

The AI was good at listening. It struggled at understanding.

The scribe correctly attributed who said what across a complex, multidisciplinary conversation, which to be clear, isa real technical achievement. But attending physicians repeatedly caught it inventing clinical content: tests and treatments that were never discussed, never ordered, never given. Every note needed a line-by-line review to catch these confabulations, which offset the time savings the tool was supposed to deliver. 

The deeper issue was context, not hallucination. The system operated outside the EHR, so notes had no access to a patient's labs, medication history, or clinical course. And a lot of what a physician knows about a case is implicit, or built from data in the record that is not  spoken aloud during rounds — which ambient audio alone can't capture.

Transcription is not clinical reasoning

Listening to a conversation and summarizing it is fundamentally different from understanding a patient's full clinical picture and reasoning through what sits in the note. A tool that only does the former will keep generating content that sounds plausible but isn't grounded in anything real. For a physician signing that note, "plausible but wrong" isn't a convenience trade-off — it has the potential to not only make the legal medical record incorrect but can also lead to real harm to the patient. 

Ambient technology works reasonably well for a single outpatient visit, where the conversation largely is the encounter. Inpatient care is different. After the initial history, much of the hospital course lives in what isn't said out loud: trending labs, medication changes, flowsheet data, a diagnosis implied by three data points but never stated in the room. Notes built only from dialogue are missing the story…and physicians are left to fill the gap.

Notes built only from dialogue aren't the complete patient story. They're a transcript.

How to spot clinical reasoning in a notewriting tool

The pilot doubles as a useful diagnostic for evaluating any AI notewriting tool. A few questions worth asking:

1. Can it show its source? Real clinical reasoning points to the lab value, medication, or vital sign behind a suggestion. A transcription tool typically can only point back to something said out loud. And if nothing was said, it has nothing to point to, which is exactly how some confabulations happen.
2. Does it know the patient's history, or just today's conversation? Check whether a note reflects trends — a lab drifting over days, a recently adjusted medication. If it only reflects the last hour in the room, it's summarizing a conversation rather than reasoning through the complete patient case.
3. Does review get faster over time, or stay just as slow? A tool generating content faster than it generates trust means review never gets lighter. Real clinical grounding should mean less time hunting for invented content.
4. Does it surface reasoning that wasn't said out loud? Clinical judgment often lives on a flowsheet before it's spoken in rounds. A tool that only captures what was verbalized will always miss those parts of a case.
5. Does it keep the clinician in the loop, or ask them to sign off on autopilot? Trustworthy tools show their evidence and let the physician confirm it, rather than handing over a finished note with nothing left to question. 

A tool that answers all five with a straight "yes" is doing something fundamentally different than one that's just transcribing well.

Where SmarterNotes fits

SmarterNotes is built to answer that standard — notewriting as clinical AI, not ambient transcription. It draws on real EHR data and 21M+ clinically-validated hospital encounters, with algorithms designed by physician-data-scientists that capture over 34,000 diagnoses and procedures, grounded in the same kind of data physicians already trust, not a transcription engine with a bigger vocabulary.

That grounding is why review becomes a quick confirmation instead of a forensic search for hallucinations — and why the time saved shows up well beyond the note itself. 

By the numbers — Up to 76% fewer pajama time nights. 10% faster discharge summaries. 10K+ diagnoses and procedures captured. 21M+ clinically-validated hospital encounters.

On fact-checking, this is where clinical judgment still matters most. SmarterNotes shows evidence alongside every recommendation rather than handing over a finished note to accept as-is. By design, that keeps the physician in the loop — reviewing and confirming, not just signing off. AI can surface a recommendation, but it still takes a clinician's judgment to catch what's off before it hits the chart. 

SmarterNotes is built to make that judgment call fast and easy, not to replace it. None of this makes SmarterNotes a scribe with better transcription. It's a fundamentally different kind of tool — built on clinical data instead of captured dialogue, from the first draft.

Physicians shouldn't have to think in ICD-10 codes, and they shouldn't have to spend more time assembling the record than applying their judgment. The lesson from the pediatric ICU pilot isn't that ambient AI is a dead end, but that listening was never the hard part. Reasoning through a patient's full clinical picture always was.

Stop simply transcribing. Start reasoning.

‍