The note writes itself. Now what?
Ambient documentation is the first AI most clinicians will meet at work, and it arrives already switched on. This is the orientation: what it genuinely does, where it fails, and what stays your responsibility.
01 What actually happens
The mechanics are simpler than the marketing. With consent obtained and the microphone on, the system listens to the conversation between clinician and patient. It handles multiple speakers, and it handles more than one language. When the encounter ends, it produces a structured clinical note in the format your specialty expects, and it can capture orders discussed during the visit.
You then read it, fix what is wrong, and sign it. That last sentence is the entire lesson, and we will spend a whole lesson on it later.
Ambient documentation does not make clinical decisions. It converts a conversation you were already having into a document you were going to write anyway. The value is real and it is narrow: it gives you back the part of the day spent typing rather than the part spent thinking.
02 Why this one landed when other health AI didn't
Plenty of clinical AI has been announced with fanfare and quietly withdrawn. Ambient documentation stuck, and it is worth understanding why, because the reason tells you where it is trustworthy.
It solves a problem clinicians actually complain about, unprompted, in every survey: documentation burden. It does not require changing clinical judgement, only clerical work. It fails visibly rather than silently — a wrong note is a wrong note you can see, unlike a subtly wrong risk score. And the human review step was designed in from the start rather than bolted on after a regulator asked.
Microsoft says more than 100,000 clinicians use it as part of daily practice, and organisations including Mount Sinai Health System and Tampa General Hospital have deployed it. Those are the vendor's own figures and named references — useful as evidence that this is past the pilot stage, and not a substitute for measuring what happens in your own department.
03 What it grew into
The 2026 version is no longer only a scribe. At HIMSS in March 2026 Microsoft positioned it as a unified clinical assistant, which means a few concrete additions:
| Capability | What it means day to day |
|---|---|
| Ambient note generation | The original function — conversation to structured, specialty-specific note |
| Direct dictation | For documentation outside a recorded encounter, the Dragon dictation people already knew |
| Order capture | Orders discussed at the point of care are captured rather than re-entered from memory |
| Role-specific workflows | Purpose-built flows for nurses and radiologists, not physician templates with the labels changed |
| Microsoft 365 Copilot integration | Pulls work context — schedule, email, documents — into the clinical workflow |
| Agents and marketplace | Extensions from partners for specific clinical and administrative tasks |
Departments that adopt ambient documentation successfully and departments that abandon it after a few months use the same software. What most reliably separates them?
Have an answer? Open this.
The most common difference is what happened in the first two weeks. Early notes need real editing — the system has not learned a clinician's patterns and the clinician has not learned how to speak in a room with a microphone in it. Clinicians who were told to expect that, and given a review method, push through. Clinicians who were told it would save them time and instead found themselves rewriting notes conclude it does not work and stop.
The technology is identical. The expectation-setting is not. Which is a training problem, and it is why this track exists.
OpenAI joined the chart. ChatGPT for Healthcare can now connect to Epic, so a clinician can ask what changed since a patient's last visit, which labs to review before the appointment, or whether medications or specialist recommendations moved — with pointers back to the supporting chart entries. A separate Healthcare Public Data plugin gives structured access to nine official sources (ClinicalTrials.gov, CMS Coverage, RxNorm, DailyMed, PubMed among them). Same rules as Dragon: it is provisioned by your organisation under a Business Associate Agreement; the EHR integration is not available on individual accounts, and features vary by deployment. OpenAI's own figure — physicians rated 99.1% of 4,363 responses safe across 27 clinical use cases — is the vendor's evaluation, not an independent one. Pilot partners named include UCSF Health and AdventHealth.
04 What does not change
Three things survive entirely, and every serious clinical governance conversation comes back to them.
The note is yours. You sign it. It is the legal record. If it says something that did not happen, that is your signature under it. No documentation tool has ever shifted that, and this one does not either.
Consent is not optional or improvised. Recording a patient encounter has legal and ethical requirements that vary by jurisdiction. Your organisation has a process. Use it.
Clinical judgement is untouched. The tool documents what was said. It does not decide what should happen, and treating a fluent note as clinical validation is a category error.
Before the next lesson
Find out three things about your own setup: your organisation's patient consent process for recorded encounters, whether your notes are reviewed by anyone other than you before signing, and who to tell when the system gets something wrong. The third one is the one nobody knows, and it is how the system gets better.