AI Scribes as Medical Devices: UK MHRA Guidance
Most ambient AI scribes are marketed as administrative tools: they transcribe a conversation, draft a summary, save a clinician time. New 2026 guidance from the UK's Medicines and Healthcare products Regulatory Agency (MHRA) draws a hard, specific line between scribes that stay in that lane and scribes that quietly cross into regulated medical device territory, often through a single feature the manufacturer thought was harmless. This guidance applies to products placed on the Great Britain market (England, Wales, Scotland); Northern Ireland follows EU MDR under the Windsor Framework. If your product touches the UK market at all, this distinction has real regulatory consequences.
- MHRA's core test is intended purpose, not the technology used. A scribe that only transcribes or summarizes for clinician review is not a medical device, regardless of how sophisticated the AI is.
- The moment a scribe suggests diagnoses, recommends treatment, or acts on clinical information without clinician review, it becomes a medical device and must meet medical device regulations.
- A disclaimer does not protect a product if diagnostic or treatment-related features are still visible in the interface or marketing elsewhere. MHRA calls warnings the weakest form of risk mitigation.
- Products can cross into medical device territory after launch, simply by adding a feature. Manufacturers are expected to reassess classification with every meaningful product change, not just at initial release.
- MHRA gives 9 concrete illustrative examples, 5 confirmed non-devices, 4 confirmed devices, and the difference between them is often one specific feature, not the product's overall category.
The Core Test: Intended Purpose, Not Technology
MHRA is explicit that the underlying AI technology, including GenAI and large language models, does not by itself determine whether a product is a medical device. What matters is intended purpose: what the manufacturer's labelling, instructions for use, and promotional materials claim the product does. Two AI scribes built on identical underlying technology can land on opposite sides of the medical device line, purely based on what one is marketed to do that the other isn't.
A specific trap worth naming: general disclaimers like "this product is not for diagnosis" do not exempt a product if medical claims are made or implied elsewhere, in the interface, sales materials, or advertising. MHRA calls warnings the lowest level of risk reduction. Manufacturers are expected to design the risk out first, not label around it.
9 Real Examples, Straight From the Guidance
MHRA's guidance includes 9 illustrative examples, split evenly enough to show how close some of these calls actually are.
An ambient scribe that provides a raw transcript of a clinical conversation for the clinician to review and edit, with no clinical suggestions of any kind.
An ambient scribe that summarizes the conversation, rather than transcribing it verbatim, again purely for clinician review before it's saved to the record.
A product that organizes information from a transcript into structured fields, like problem lists or medication lists, for the clinician to review and confirm, without deriving new clinical information.
A product that suggests clinical billing codes based on terms explicitly mentioned in the conversation, matching stated language to reference codes rather than inferring anything.
A product that drafts a discharge summary or letter using conversation data plus existing information from the patient's record, for clinician review and edit before use.
The one worth studying closely. A summarization tool that offers an optional "generated insights" feature suggesting possible diagnoses or treatment options, with a disclaimer that the feature "is not intended to replace traditional methods of diagnosis or treatment." MHRA's ruling: the disclaimer doesn't matter, the feature itself is diagnostic, and the product is classified as a medical device regardless of how the feature is labelled.
A scribe that is functionally identical to Example 2, pure summarization, but whose manufacturer markets it with the claim that it "guides diagnosis and treatment planning and improves patient outcomes." The marketing claim alone is enough to make it a medical device, even though the product's actual functionality never changed.
An AI agent that finalizes and saves transcripts to the patient record without clinician review, then autonomously determines follow-up tests are needed and places the orders without clinician input. This is the clearest case: the product is taking clinical action a professional would normally take or oversee.
A report-drafting tool that also analyzes conversation data and test results to generate suggested diagnoses, where the reasoning behind the suggested diagnosis may not be shown at all, meaning the clinician is expected to rely on the output rather than their own judgment.
Why Example 6 Is the One to Actually Worry About
Examples 8 and 9 are extreme enough that most teams would recognize the risk immediately. Example 6 is the dangerous one, because it describes a product that was designed correctly, marketed responsibly, and even added an explicit warning, and still got classified as a medical device. The lesson isn't "add better warnings." It's that an optional feature offering diagnostic or treatment content changes the product's regulatory status the moment it exists in the interface, whether or not a user ever clicks it.
Products Can Cross the Line After Launch
MHRA notes explicitly that many AVT products are built on flexible, general-purpose GenAI technology that makes it easy for manufacturers to add capabilities well beyond a product's original administrative scope. A scribe that launches as a pure transcription tool and later adds a diagnostic-suggestion feature doesn't get to keep its original "not a medical device" status. Manufacturers are expected to reassess classification with every meaningful product change, not treat the initial launch determination as permanent.
If It Is a Medical Device: What Classification Looks Like
Once a product qualifies as a medical device under UK MDR, classification follows specific rules for active software devices. The relevant ones for AVT products:
| Device Behavior | UK MDR Classification |
|---|---|
| Allows direct diagnosis, or provides decisive information used to make a diagnosis | Class IIa |
| Monitors vital physiological parameters where variation could cause immediate danger (e.g. cardiac performance, respiration) | Class IIb |
| All other active software devices not meeting the above | Class I |
Class IIa and above require certification by an approved body before the product can legally be placed on the GB market. MHRA's guidance also clarifies that "decisive information for diagnosis" can cover a broad range of outputs, including a scribe that estimates the probability of a condition based on a clinical interaction, if that output is intended to be relied on for treatment decisions.
Also named explicitly in the guidance: hallucination is called out as a known GenAI hazard with potentially broad impact specifically in AVT products. Manufacturers of GenAI-based scribes are expected to mitigate this risk in alignment with the product's stated intended purpose, not treat it as an acceptable side effect of the technology.
This regulatory question sits alongside, but is distinct from, the privacy and data-handling questions we've covered for ambient scribes generally. See our companion pieces on how ambient AI scribes handle recorded audio, what an AI medical scribe actually is, and what real hospital readiness for AI scribe adoption looks like.
The NHS England Side: Practical Implementation, Not Just Classification
MHRA's guidance answers one question: is this product a medical device? A companion document, NHS England's guidance on adopting AI-enabled ambient scribing products, answers a different one: how do you actually deploy one responsibly, regardless of its classification? A few specifics worth knowing if HosTalky or any comparable platform is evaluated against it:
- Pilots have a hard time limit. Running a "pilot" or proof of concept does not exempt a supplier from compliance requirements. NHS England explicitly caps this: pilots can continue only if they are time-limited to four months and aren't being used to bypass compliance obligations.
- Accent and dialect bias is named directly as a risk. The guidance states that AVTs may have varying success with different accents or regional dialects, and for speakers using English as a second language or affected by speech disorders. Deploying organisations are told to confirm the product works well across these groups and monitor whether any patient group hesitates to seek care because of it. This is a specific, named subtype of the broader AI bias problem we've covered in real-world examples of AI bias in healthcare.
- Patients have a right to not be subject to fully automated decisions under UK GDPR Article 22, where the outcome has a legal or similarly significant effect on them, relevant directly to Example 8 and 9's autonomous-action scribes above.
- Data and cybersecurity risk is treated as a distinct category from clinical/safety risk, including risks specific to externally hosted AI models and the "black-box" nature of third-party AI systems making privacy impact assessments harder to complete. This overlaps with concerns we've raised about AI-built healthcare apps and their security posture generally.
An AI Scribe Built to Stay in Its Lane
HosTalky's AI Scribe is designed around documentation and clinician review, not diagnostic suggestion, keeping the clinical judgment where it belongs.
See How AI Scribe WorksThe Bottom Line
The difference between a pure documentation tool and a regulated medical device usually isn't the underlying AI, it's one feature: a diagnostic suggestion, a treatment recommendation, an autonomous action taken without clinician review. That feature can exist as an "optional" toggle, carry a disclaimer, and still trigger medical device classification. For any team building or evaluating an ambient scribe with UK market exposure, the practical takeaway is to audit every feature against this specific test, not just the product's overall stated purpose, and to re-run that audit every time a new feature ships. If you're still comparing whether an AI scribe is worth adopting at all, our breakdown of AI scribe vs. manual documentation covers the time and accuracy trade-offs independent of this regulatory question.
FAQs
Is an ambient AI scribe a medical device?
Does adding a disclaimer protect an AI scribe from being classified as a medical device?
Can an AI scribe become a medical device after it launches?
What classification would a diagnostic AI scribe receive under UK regulations?
Sources and References
- Medicines and Healthcare products Regulatory Agency (MHRA). (2026). Guidance on Ambient Voice Technology-Enabled Products: Device Qualification and Classification, Version 1.0. gov.uk
- NHS England. (2026). Guidance on the Use of AI-Enabled Ambient Scribing Products in Health and Care Settings, Version 3. england.nhs.uk
- MHRA. Borderline Products: How to Tell if Your Product Is a Medical Device. gov.uk
