The situation
A provider group running 34 clinics across four states had a retention problem that presented as a technology problem. Physician turnover had reached nineteen percent annually, and the exit interviews kept naming the same thing: documentation burden.
A time study they had commissioned put the average physician at two hours and forty-five minutes per day on notes, most of it after the last patient had left. Their electronic health record was not going anywhere, for reasons that were contractual rather than technical. Whatever we built had to live alongside it.
They had already tried two things. A voice dictation product that produced transcripts requiring almost as much editing as typing from scratch. And a template library, which helped the first week and then became another list to scroll past.
What we found in the architecture sprint
We spent the first week with clinicians rather than with the codebase. Three findings changed the shape of the project.
The bottleneck was structure, not typing. Physicians were not slow at typing. They were slow at deciding what belonged in which section, at chasing values from three different screens, and at reformatting the same information for the note, the billing code and the referral letter. The same facts were being entered up to three times in different shapes.
Nobody trusted a system that wrote the note alone. Every clinician we spoke to said some version of: I will sign it, so I have to read every word anyway. Any design where the model produced a finished note and the physician approved it in bulk was going to be rejected, and rightly.
The existing integration was the real constraint. Their EHR exposed a FHIR API with rate limits that made naive polling impossible, and a subset of the data lived only in an HL7 v2 feed. That shaped the architecture more than any product decision.
What we built
A structured capture layer during the encounter. Rather than a blank note, clinicians work through a compact structured form driven by the visit type, with fields prefilled from the FHIR record. Vitals, medications, allergies and recent results arrive already populated. The physician confirms or corrects rather than retrieves and retypes.
This alone accounted for a little over half the total time saving, and it involved no machine learning at all. It is worth naming, because the AI is what gets attention and the boring integration work is what did most of the job.
Model assisted drafting with mandatory review. From the structured capture and the ambient transcript, the system drafts a narrative note. The draft is always presented in an editing view, never as a finished document. Every generated statement is linked back to the transcript segment or structured field it came from, so a physician can check any sentence in one click.
Three design decisions made this acceptable to clinicians:
- The draft never invents a clinical finding. Generation is constrained to content that traces to a source, and unsupported statements are dropped rather than smoothed over.
- Low confidence sections are flagged rather than hidden, which means the physician's attention goes where it is needed instead of being spread evenly.
- Nothing is signed automatically. Ever. The signature is a deliberate act on a note the clinician has read.
An evaluation harness before deployment. We built a golden dataset of 240 encounters with clinician written reference notes, and measured factual accuracy, omission of clinically significant findings, and hallucinated content on every model or prompt change. Omission of a significant finding was treated as a blocking failure, not a score. That harness ran in continuous integration for the whole build and still runs weekly against production samples.
HIPAA aligned infrastructure from the first commit. Dedicated AWS account, all data encrypted at rest with customer managed keys, audit log on every record access, no protected health information in application logs, and an enterprise model tier with a business associate agreement and no training on inputs. Infrastructure defined in Terraform so the compliance posture is reviewable as code rather than as a claim.
What was harder than expected
Specialty variation. The first pilot was in primary care and went well. Rolling into orthopaedics and behavioural health exposed how much note structure varies by specialty. Behavioural health in particular has documentation conventions where the same words carry different weight. We had to build per specialty configuration and rebuild the golden dataset three times, which added about six weeks.
The HL7 v2 feed. Roughly fifteen percent of the data we needed was only available through a nightly HL7 v2 batch, which meant certain fields could be up to a day stale. We could not fix the source system. Instead we made staleness visible in the interface, showing the timestamp on any value that came from the batch feed, so clinicians knew when to verify. Displaying the limitation honestly turned out to be more trusted than hiding it.
Trust took longer than accuracy. By month four the system was measurably accurate. Adoption was still under thirty percent. What moved it was not a better model but a change in onboarding: pairing each new clinician with a colleague already using it for two sessions. Adoption reached eighty one percent within six weeks of that change. The lesson we took was that a clinical tool is adopted socially, not technically.
Results after seven months in production
Documentation time fell from two hours forty-five minutes to one hour thirty-four minutes per physician per day, a forty-three percent reduction, measured by the same time study methodology used for the baseline.
Ninety-six percent of generated drafts were accepted with light editing, defined as fewer than fifteen percent of characters changed. Four percent were substantially rewritten or discarded, and every one of those went into the evaluation dataset.
Note completion moved forward in the day. Before, sixty-two percent of notes were signed after 18:00. After, that figure is nineteen percent.
Physician turnover in the twelve months following full rollout was eleven percent, down from nineteen. Documentation is not the only variable and we would not claim the whole difference, but it was the reason clinicians gave.
There have been no reportable privacy incidents. The audit trail has been exercised twice by internal compliance review and satisfied both.
What we would do differently
We would build the per specialty configuration from the start rather than generalising from primary care. The rework cost about six weeks and was foreseeable.
We would also start the social onboarding programme in the pilot rather than in month five. We treated adoption as a downstream consequence of quality, and it was not. It was its own workstream and it deserved a plan from day one.