What do AI scribes actually do, and what should a practice check before using one?

A physician friend described her evenings to me as “the second shift.” Clinic ended at five and the notes ended at nine. That is the problem ambient AI scribes are sold to fix, and in the last year the evidence has finally moved past vendor case studies to randomized trials. The results are more modest than the marketing, and more useful.

What does an AI scribe actually do during a visit?

It listens to the conversation between clinician and patient, usually through a phone or computer in the room, turns the audio into text, and drafts a structured visit note. The clinician then reviews, edits and signs the note in the record. The scribe writes a first draft. It does not examine, diagnose or decide anything.

That last line matters for how you set it up. The signed note is the clinician’s note, whatever drafted it. Any workflow that lets a draft reach the chart without a person reading it has turned a writing aid into an unsupervised author of the medical record.

Do AI scribes save clinicians time?

Some, and less than the brochures suggest. In a UCLA trial published in NEJM AI in 2025, 238 outpatient physicians were randomized to one of two scribes or usual care. One scribe reduced time in notes by 9.5% against control. The other showed no significant change. Both groups reported some improvement on burnout and task load measures.

A second randomized trial, run across ambulatory clinics in two states and also published in NEJM AI in 2025, followed 66 practitioners over 24 weeks. Time spent on notes fell by about 0.36 hours a day, and work exhaustion scores fell. Professional fulfillment did not improve significantly. Neither trial is the last word, and both were short.

The two-state trial also had professional coders review billing codes and scored note quality, and reported that documentation time fell without hurting either. That is one trial in one health system. It is still the kind of check worth copying in your own pilot: have someone score a sample of AI-drafted notes against your own standard before you widen use.

How accurate are AI scribe notes?

Accurate enough to be useful and not accurate enough to skip review. In the UCLA trial, clinicians rated clinically significant inaccuracies as happening “occasionally” with both products, and the authors concluded that those inaccuracies require ongoing vigilance. Plan on every note being read before it is signed, and on someone checking a sample in detail.

Practically, that means three things. Tell clinicians which errors to watch for in your first weeks, such as a medication dose, a laterality or something the patient said in passing that ended up stated as fact. Keep a simple log of corrections. And when the same error keeps coming back, raise it with the vendor, because it tells you where the tool is weak in your kind of visit.

Do patients have to consent to an AI scribe?

Plan on asking every patient, and check your state’s recording law. The UCLA trial ran in California, and the authors say that under the state’s two-party consent law for audio recording, physicians were told to get verbal consent from everyone present and document it. The two-state trial informed patients at check-in and again verbally, and reports that 22 declined recording.

Give patients an easy way to say no that costs them nothing, and make sure staff know how to switch the scribe off for that visit. Some patients will ask what happens to the recording. Your front desk should be able to answer in one sentence, which means you need to know the answer first.

Put a short notice at check-in as well, in plain words: what the scribe does, that a clinician reviews every note, and that the patient can say no without any effect on their care. A patient who read that at the desk arrives in the exam room already knowing the answer to the first question they would have asked.

What should a practice check before piloting an AI scribe?

Start with the paperwork. A scribe vendor hears and stores patient conversations, so HHS treats it as a business associate and a signed agreement comes before the first recorded visit. Then ask where audio and transcripts are stored, how long they are kept, whether they are used to train models, and whether notes flow into your record system cleanly.

1. Get the BAA signed and update your risk analysis for the new tool.
2. Write the consent script and the opt-out step, and train the front desk on both.
3. Pick three to five clinicians who want to try it, not the ones who were told to.
4. Measure before and after: time in notes and after-hours charting from your record system, plus a short survey.
5. Log every correction clinicians make in the first month, and review the log with the vendor.
6. Decide at a set date whether to widen, change or stop, using those numbers.

Who will not get much from an AI scribe?

The trials hint at it. In the UCLA study the two scribes were used in 33.5% and 29.5% of visits, which tells you many clinicians left them off even when they had access. A clinician who already finishes notes during the day, or whose visits are short and templated, has less to gain. And if the real bottleneck is inbox messages or prior authorizations, a scribe will not touch it. Look at where the evening hours actually go before you buy.