MHRA Draws the Line on AI Scribes as Medical Devices

MHRA and NHS England published joint guidance on 29 July 2026 drawing a line that NHS trusts, and the companies selling into them, have been asking for: when does an ambient voice technology tool — an AI “scribe” that listens to a clinical conversation and produces a transcript or summary — count as a medical device? The guidance, titled “Ambient voice technology-enabled products,” does not change the underlying law. It applies the UK's existing medical device qualification rules to a product category that has spread across the NHS faster than the regulatory clarity around it.

Why AI scribes needed a line drawn

Ambient voice technology — software built on large language models that listens to a consultation and turns it into a structured note — has become one of the fastest-adopted AI categories in the NHS, sold as a way to give clinicians back the time they currently spend typing. The regulatory question it raises is not new in kind: software has qualified as a medical device based on its intended purpose since long before generative AI existed. What is new is the volume of products making claims that sit close to the line, and the pressure that puts on procurement teams and information governance boards to make a classification call themselves, product by product, without a clear test to apply.

The guidance gives them that test. On one side: a tool intended solely for transcription, summarising a clinical conversation, drafting a letter, or suggesting a clinical code — provided a clinician reviews the output before acting on it. Those functions, per the guidance, do not qualify a product as a medical device. On the other side: a tool intended to support diagnosis, treatment, or disease prevention, or one that takes automated action — the guidance's example is placing an order — without a clinician reviewing it first. Either of those crosses into UK MDR territory, with the obligations that follow: a manufacturer has to demonstrate safety and performance, not just usefulness.

The clinician stays the backstop

The guidance is explicit that the classification test does not shift where responsibility sits. Whether or not a given AVT product counts as a medical device, a clinician is still expected to review and verify the transcript or summary it produces before it becomes part of a patient's record or informs their care. Classification decides which regulatory obligations attach to the software; it does not decide who answers for what the software gets wrong. That distinction matters for procurement teams evaluating AVT tools: a product qualifying as "not a device" is not a product free of clinical governance requirements — oversight, staff training, and monitoring for functionality changes that could alter a product's status all remain expected of the NHS organisation deploying it.

Part of a wider AI-in-health push

MHRA and NHS England frame the guidance as supporting the NHS's stated ambition to become, in their words, the most AI-enabled health system in the world — and as feeding into the work of the National Commission into the Regulation of AI in Healthcare, a non-statutory advisory body working with MHRA, NHS England, and other stakeholders on the broader regulatory framework AI-in-health products will eventually sit under. For now, this guidance changes nothing about the underlying law. It is MHRA telling the market, in plain terms, how that law already applies to a product category regulators did not have to imagine five years ago and now cannot avoid.