10decoders
The Note Looks Finished, and That Is the Problem
A primary care physician finishes her eleventh visit of the morning. The ambient scribe has already drafted the note: history, exam, assessment and plan, in clean paragraphs that read better than anything she would have typed between rooms. She skims it, corrects a medication dose, and signs. What she misses is that the patient said the chest tightness comes on when he climbs stairs, and the note records no exertional symptoms. Nothing about the draft looked wrong, and that kind of error is the one health systems need to plan for.
The case for these tools is real. In a JAMA Network Open study published in October 2025, Olson and colleagues followed 263 clinicians across six health systems and found burnout fell from 51.9% to 38.8% after 30 days of ambient scribe use. A 2025 survey of 43 health systems in JAMIA, led by Eric Poon at Duke Health, found every respondent had adoption work under way for ambient notes, the only AI use case with universal uptake. Clinicians want these tools, and leadership has good reasons to keep them.
Note quality is a separate question from burnout, and it gets far less measurement. A fluent draft changes how people read. Writing a note from scratch forces a clinician to recall each element. Reviewing a polished draft invites skimming for obvious mistakes. The errors that matter most in published evaluations are the ones skimming misses: a dropped negation, a symptom on the wrong side, a plan item said aloud that never reached the text. Most go-live plans treat the clinician's signature as the control. It is a control, but almost nobody measures how well it works.
A draft that reads well gets reviewed less carefully, and the errors that matter most are the ones that read well.
Which Scribe Errors Do the Most Damage
Published studies sort scribe errors into a handful of recurring types. The severity column below is our working triage template from healthcare delivery work, informed by the npj Digital Medicine and JMIR findings above. It is not a clinical consensus, and your clinical safety lead should own the final ratings.
| Error type | What it looks like in the note | Severity |
|---|---|---|
| Dropped or flipped negation | “Denies chest pain” becomes “chest pain,” or the reverse. Hard to see inside a fluent sentence, and it can change the differential | Critical |
| Fabricated finding or history | An exam element, medication or history item that was never discussed appears as documented fact | Critical |
| Omitted plan item or instruction | A referral, follow-up interval or medication change spoken in the room never reaches the plan | High |
| Wrong attribution | A parent's history recorded as the patient's, the wrong laterality, or a statement assigned to the wrong speaker | High |
| Carried-over spoken error | A misstatement by the patient or an interpreter is transcribed faithfully and reads as confirmed | Moderate |
| Verbosity and template drift | Notes grow longer, repeat content across sections, or stop matching the specialty template | Lower |
Not sure where your ambient scribe accuracy gaps are?
10decoders builds evaluation and monitoring pipelines for healthcare AI inside HIPAA-aligned environments. We help clinical informatics teams sample notes, score them against the encounter, and turn the results into a number leadership can track.
Book a Free AI Assessment →Why the Vendor's Accuracy Number Is Not Your Accuracy Number
Published error rates are hard to compare because they count different things. Asgari and colleagues measured per sentence on mock consultations. Biro and colleagues measured per note on real encounters. Both results can be true at once: combined, a hallucination rate near 1.5% and an omission rate near 3.5% per sentence mean a 30-sentence note will usually contain at least one error, if errors are independent, which they may not be. When a vendor quotes an accuracy figure, ask what the denominator is, which encounters it came from, and whether anyone corrected the output before it was scored.
Your own rate depends on conditions the vendor does not control: specialty vocabulary, the languages and accents in your patient population, exam room acoustics, how many people speak during the visit, and the note templates your EHR team configured. A behavioral health intake, a pediatric visit where a parent answers most questions, and an orthopedic follow-up will each produce a different error profile. A UK review published by Healthwatch England in late August 2026 collected examples from routine use, including a dropped negation and a fabricated medication.
The rate also moves over time. Vendors update models and prompts on their own release schedule, and a change that improves one specialty can quietly regress another. A pre-go-live validation tells you how the tool performed on the day you tested it. It says nothing about how it performs six months later, after three updates you may not have been told about.
Accuracy Checked Once, Before Go-Live
A pilot group reviewed sample notes and approved the tool. No ongoing sampling exists. The signing clinician is the only control, and nobody tracks how often drafts are edited.
Clinicians Can Flag Bad Notes
A flag button or ticket route exists and informatics reviews what comes in. The data only covers errors clinicians happened to notice, so it skews toward obvious mistakes and misses the fluent ones.
Sampled, Scored and Tracked by Specialty
A fixed sample of notes per specialty is scored against the encounter every month. Error rates by type go to the clinical safety committee, and every vendor update triggers a re-score.
Checklist: What Post-Go-Live Scribe Monitoring Should Include
Use this list to test your current program. Each item should point to a named document, report or owner.
The signature on the note is a control. Measure it like one, or stop counting it as one.
What to do this week
01 Score twenty notes from your busiest specialty
Pick the specialty with the highest ambient scribe volume and pull twenty recent encounters where the transcript is still available. Have two reviewers, ideally a clinician and a clinical documentation specialist, score each draft against the transcript using four error types and a major or minor grade. Within a week you will have a baseline error rate per note and per type, and many programs running these tools at scale have never produced one.
02 Ask your vendor three questions in writing
What is the denominator behind your quoted accuracy figure, per sentence, per element or per note? Which encounters produced it, and were they simulated or live? How are model and prompt updates communicated, and can you receive release notes before they ship? Keep the answers on file. Vague replies are useful information for your next contract review.
03 Pull edit rates from the EHR
Ask your EHR analytics team for the share of AI-drafted notes signed with no edits, split by specialty and by note section. Look hardest at assessment and plan sections in high-acuity visits that are signed untouched. A low edit rate can mean the drafts are good, or it can mean nobody is reading them. Only sampling tells you which.
04 Name an owner and write a pause rule
Assign one clinical lead to own scribe accuracy, and agree a threshold for major errors in the monthly sample that triggers mandatory full reads or a pause for that specialty. Record the rule in your AI governance minutes, so the decision to act does not depend on someone finding the time to argue for it after an incident.
Let 10decoders Measure Your Ambient Scribe Accuracy
We set up a scored sampling program for one or two specialties, build the error taxonomy with your clinical informatics team, connect edit-rate data from your EHR, and hand over a monthly accuracy report with clear thresholds. The work runs inside your HIPAA controls and alongside your existing vendor contract.