5 August 2026 · Medows
AI Scribe or Medical Device? The UK Decides
The MHRA just clarified when an AI scribe counts as a medical device in the NHS, and what it means for the doctors relying on one.
A form with two boxes
Picture a GP in Leeds, mic on, ambient scribe listening. It writes the note. It suggests a code. She reads it, edits a line, signs off. Nothing about that visit needed a CE mark, not according to the MHRA.
Now picture a version of the same tool that also flags "possible UTI" before she's said a word out loud. That one needs a CE mark. Same room, same conversation, different regulatory universe.
On 29 July 2026, the UK's Medicines and Healthcare products Regulatory Agency tried to draw that line clearly for the first time. The guidance, developed with NHS England, splits ambient voice technology into two boxes.
Box one: administrative
Tools that transcribe a conversation, summarise it, draft a letter, or suggest a clinical code for a human to check, stay out of medical device regulation. They're treated like a very good stenographer: useful, unregulated, and entirely dependent on the clinician reading what they wrote.
Box two: clinical
Tools meant to support diagnosis or treatment, or that take an action on their own (placing an order without a clinician confirming it first, for instance), cross into Software as a Medical Device territory. That means the full weight of UK medical device law: safety evidence, post-market surveillance, a CE or UKCA mark.
MHRA chief executive Lawrence Tallon put it simply: "This guidance aims to provide clarity on where the regulatory line sits for this high-growth technology." NHS England's chief digital officer, Rob Thompson, added that "AI notetaking tools are already showing real promise in reducing administrative burden on frontline staff," while noting that trusts still need governance, training, and procurement checks in place before rollout.
Why this matters on the ward, not just in a compliance memo
The interesting part isn't the boxes. It's what decides which box a tool lands in: the manufacturer's stated intent, not strictly what the software does under the hood. A tool labelled "for transcription only" that also happens to be very good at pattern-matching symptoms is, on paper, still box one. The line is drawn by the label, and the clinician using the tool is the one left standing in the gap between the label and the behaviour.
That gap is exactly where "AI helped me write this note" quietly turns into "AI helped me decide this diagnosis," without anyone updating the paperwork. A junior doctor twelve hours into a shift isn't going to pause and ask which regulatory box the tool in front of her falls into. She's going to trust what's on the screen, because it's fast, and because it's usually right.
Which is the actual case for building AI for clinicians as a workspace with a paper trail, not a black box with a transcript. Every suggestion should be traceable to where it came from in the chart. Every action a clinician didn't explicitly take shouldn't quietly happen anyway. Regulation is drawing that line from the outside, one guidance document at a time. The tools worth trusting on a ward round should already be drawing it from the inside, before a regulator ever has to ask.
Sources
Medows is a clinical AI workspace for the doctor on rounds. Learn more or write to alapan@medows.ai / alapanx@gmail.com.