UK Draws a Regulatory Line for AI Medical Scribes: Documentation Alone Is Not a Medical Device, but Entering Diagnosis and Treatment Brings Regulation

AI operating quietly in consultation rooms may begin simply by organizing clinician-patient conversations into medical records. Once it starts recommending diagnoses or treatments, or directly triggering healthcare processes, its role shifts from an administrative assistant to medical software that may influence care decisions. The UK Medicines and Healthcare products Regulatory Agency (MHRA) recently clarified this boundary, setting out regulatory criteria for AI medical scribes rapidly entering the National Health Service (NHS).

According to the MHRA, if a product’s intended purpose is limited to verbatim transcription, summarizing clinical conversations, drafting letters, or suggesting clinical codes for review by healthcare professionals, it is not considered a medical device under the current framework. By contrast, if the tool is used to support the diagnosis, treatment, or prevention of disease, or automatically orders tests or submits clinical orders without confirmation by a clinician, it must comply with the corresponding medical device safety and performance requirements.

This clarification does not amend the law, but explains how existing rules apply to ambient voice technology. Classification depends not only on the model used by a product, but also on the “intended purpose” defined by the manufacturer in its instructions for use, labeling, and promotional materials. Consequently, two seemingly similar AI scribes may follow entirely different regulatory pathways because their outputs and subsequent actions differ.

The boundary may also shift with software updates. A system that initially generates only summaries may need to be reassessed if treatment recommendations, omission alerts, or automated workflows are added. The flexibility of generative AI further complicates the issue: even if a supplier positions a product as an administrative tool, users may still use free-text instructions to ask it to comment on diagnosis and treatment or propose next steps. Product guardrails, actual capabilities, and how the product is used therefore cannot be judged by its name alone.

Not being classified as a medical device does not mean safety governance can be waived. NHS England requires adopting organizations to verify a product’s regulatory status themselves rather than relying solely on supplier statements. Clinicians must review outputs before they enter medical records or trigger subsequent work, while healthcare institutions should establish safety cases, risk records, staff training, performance monitoring, and incident-reporting mechanisms. The MHRA has also explicitly stated that healthcare professionals’ responsibility to check AI-generated transcripts and summaries remains unchanged.

Practical risks are often hidden in everyday language. NHS guidance warns that systems may misunderstand clinical terminology, abbreviations, rapid speech, accents, dialects, or speech and language impairments, and may record a patient’s hypothetical statement as a confirmed diagnosis. Missing information, overreliance on automated output, and poor integration with electronic health records may all carry small errors into care processes. Deploying organizations must also address recordings, medical-record data retention, and cybersecurity, and assess whether different patient groups are disproportionately affected.

The materials published to date mainly provide a regulatory and governance framework. They do not present accuracy rates for specific products, clinical trial results, or validation data across populations, and therefore cannot be used to conclude that AI scribes have been proven safe and effective in real-world clinical care. The real significance of this boundary is that it reframes the question from “Does using AI make it a medical device?” to “What does it do in the care chain, and who performs the final check?” For developers and healthcare institutions, every step a function takes closer to clinical decision-making brings greater responsibilities for evidence and oversight.