The Last Mile of AI Scribes

The note is done. It is accurate, well-structured, medically appropriate. The AI listened to the encounter, understood the context, and produced documentation that would have taken the physician twenty minutes to write. That part worked.

Then the physician turns back to the screen and spends the next eight minutes clicking through the physical exam fields.

The technology delivered on its hardest promise. What remains is a different kind of problem, and it has a more specific shape than most people in this space have acknowledged.

The note quality problem is largely solved

It is worth pausing on that for a moment, because the industry spent years convinced it wasn't solvable. Early AI scribes produced output that required heavy editing. Physicians found themselves correcting more than they were saving. The skepticism was earned.

That era is ending. The current generation of ambient AI scribes produces medically accurate, contextually appropriate notes. The language is right. The structure holds. The content reflects what actually happened in the room.

The History of Present Illness captures the nuance of the conversation. The Chief Complaint reflects what the patient said. The Assessment and Plan is coherent and medically appropriate. Physical examination findings come back as organized, structured output. Laboratory results are summarized clearly. The scribe is not just generating prose – it is generating well-organized, clinically relevant content across the full note.

The problem is not the quality of that output. The problem is where it has to go.

EMRs are hybrid systems

This is the part that rarely gets discussed with enough precision.

An EMR is not a document editor. It is a structured database with a document-like interface layered on top. Some fields accept prose. Others require discrete, structured data: a checkbox, a dropdown, a numeric entry, a point-and-click selection.

Vitals. Physical examination findings. Review of systems in many systems. Procedure codes. Diagnostic codes. These fields exist in structured form for reasons that make clinical and operational sense. Structured physical exam findings can be scanned across a patient's record over time. Discrete data supports population health queries. MIPS and quality reporting requirements depend on it. Shared care environments – where multiple providers work from the same chart – rely on structured fields to surface the right information at the right moment.

The AI scribe produces well-structured output. A bullet list of physical exam findings. An organized summary of laboratory values. The physician can read it, edit it, approve it. What they cannot always do is paste it. Because the EMR is not waiting for text – it is waiting for each finding to land in its own discrete field, its own checkbox, its own dropdown. The scribe did the clinical thinking. The physician however, still needs to spend time on the EMR data entry for some sections at least.

That gap is the last mile. And it varies: which sections are free text, which require structured input, how granular the form elements are – all of this differs across EMR systems and user templates. There is no single mapping to build. There is a landscape of them.

What actually happens at the point of adoption

A physician tries an AI scribe. The HPI is excellent. The assessment is sharp. They paste it into the chart and feel the pull of something genuinely useful.

Then they look at what the scribe produced for physical exam. The findings are right. The structure is clear. And none of it can be pasted in, because the EMR is not expecting a bullet list – it is expecting a click on "lungs: clear to auscultation" and another click on "heart: regular rate and rhythm.". The time savings are real but partial. The post-visit computer time compresses but doesn't disappear.

This is not resistance to new technology. It is a clear-eyed response to a gap the technology hasn't closed yet. The note improved. The workflow didn't fully change.

The next generation of scribes needs to close this gap

Getting from AI-generated text to structured EMR fields is not a fully solved problem. It requires understanding which fields in which EMR accept which data types, mapping clinical language onto discrete values, and navigating a landscape of integration standards that varies considerably across systems.

It is also not an unsolvable problem. It is specific. It is bounded. The teams building in this space who understand the constraint precisely – who have heard enough physician feedback to know that the last mile is about structure, not prose – are the ones positioned to close it.

The note being good is necessary. It is not sufficient.

This post was written by the team at IrisNote. We build AI scribes for small practices, and we spend a lot of time listening to the physicians using them. The last mile problem is the one we are most focused on getting right.