Ambient documentation (AI scribe)

What it means

A doctor walks into the exam room and asks the patient whether a phone may listen in. At The Permanente Medical Group in Northern California, “Using the microphone on a secure smartphone, the ambient AI scribe transcribes—but doesn’t record—patient encounters” and then “uses machine learning and natural-language processing to summarize the conversation’s clinical content and produce a note documenting the visit.”¹ There, “Patient consent was obtained before using the ambient scribe during the visit.”¹ Machine learning and natural-language processing are the methods that let software learn patterns from examples and work with ordinary human language.

“Ambient” means the tool works in the background of a normal conversation. The doctor does not dictate to it. Microsoft and Nuance listed “ambient listening” among its abilities.² Think of a colleague who takes the minutes in a meeting so everyone else can talk. Someone still has to read those minutes before they go on file.

Where the word comes from

We could not find who first coined “AI scribe.” The earliest use of the “ambient” term we could source is from October 17, 2019. That day Nuance and Microsoft announced a partnership around “ambient clinical intelligence” that would power “the exam room of the future where clinical documentation writes itself.”² They planned to “introduce the technology to an initial set of physician specialties in early 2020.”²

CNBC later described “the original DAX application that Nuance launched in 2020.” It said DAX “converts verbal patient visits into clinical notes, and it sends them through a human review process to ensure they are accurate and high-quality.”³ In March 2023 a new version paired ambient AI with “OpenAI’s newest model, GPT-4,” and promised to start “automatically generating a draft of a clinical note within seconds after a patient visit.”³ Now “Physicians can make edits to the notes that DAX Express generates, and they sign off on them before they are entered into a patient’s electronic health record.”³ In CNBC’s account, the check moved from a separate review process to the doctor who signs. By 2024, “ambient artificial intelligence (AI) scribes” was the term in a medical journal’s title and text.⁴

How others define it

• American Medical Association (reporting on Permanente): the scribe “transcribes—but doesn’t record—patient encounters” and produces “a note documenting the visit.”¹
• *Tierney and colleagues, NEJM Catalyst:* “ambient artificial intelligence (AI) scribes, which use machine learning applied to conversations.”⁴
• UK Medicines and Healthcare products Regulatory Agency (MHRA): “AVT products, often referred to as AI scribes, are increasingly used across the NHS.” They “record and summarise conversations between clinicians and patients.”⁵ (AVT stands for ambient voice technology.)
• Health Services Safety Investigation Body (HSSIB), England: “a class of AI-enabled ‘ambient scribing’ tools used in healthcare.”⁶

Where they agree and differ

All four describe software that listens to a clinical conversation and turns it into written documentation. They differ on recording. The AMA’s source says the tool “doesn’t record,”¹ while the MHRA says these products “record and summarise”⁵ and UCLA Health says its tools “record patient conversations and automatically generate draft clinical notes.”⁷ For a leader, what happens to the audio is a privacy question to ask each vendor.

Regulators also draw a line. Under the MHRA’s guidance, tools for transcribing and summarizing are “not regulated as medical devices under the current framework.” Tools meant to support diagnosis or treatment “are regulated as medical devices and must meet the relevant safety and performance requirements.”⁵

The strongest view against the hype is that the benefits are real but modest and not yet proven. In a UCLA trial of “238 physicians across 14 specialties and 72,000 patient encounters,” users of one tool “reduced their documentation time by nearly 10% compared to usual care.” The other tool’s reduction was smaller and not statistically significant.⁷ Physicians said the notes “occasionally” had “clinically significant inaccuracies, most commonly omissions of information or pronoun errors,” and “One mild patient safety event was reported during the study.”⁷

In the news

On 6 August 2026, HSSIB gave notice of an investigation into ambient voice technology in hospitals.⁶ It said: “Intelligence gathered found that adoption of AVT is accelerating while the safety implications are not fully understood.” It adds, “We expect to publish an investigation report in summer 2027.”⁶ A week earlier, on 29 July 2026, the MHRA restated that “Clinicians remain responsible for reviewing and verifying AI generated transcripts.”⁵ We are fairly confident in these facts because they come from the bodies themselves. What the investigation will find is not known.

What it means for your organization

AI scribes show what happens when an agent writes the official record. At Permanente, “a tiny percentage” of notes “did contain ‘hallucinations,’” the AMA reported, meaning false information produced by the AI.¹ Our perspective, which is a hypothesis and not a verdict: this record holds up when a named person reads and signs every note (see [a person in the middle](/definitions/a-person-in-the-middle)). We are watching the HSSIB report and longer trials. We would change our view if evidence showed that sign-off becomes a habit with no real reading behind it. Meeting notes, call summaries, and case files face the same question.

One question to ask your team: When an AI drafts our records, who signs them, and how would we know if they stopped reading before signing?