The first generation of AI scribes wrote notes. The second generation is entering the clinical workflow — suggesting codes, drafting referrals, generating safety-netting advice, proposing investigations, and creating follow-up tasks. The technology is becoming more capable, more integrated, and more useful. It is also crossing a boundary with significant implications for clinical safety, regulation, and professional accountability.
The First Generation Wrote Notes
Basic scribing was straightforward: listen, transcribe, summarise into a structured note. The value: less typing, more eye contact, faster documentation. The clinical risk was manageable — a transcription error produces an imperfect narrative that the clinician catches during review.
The Second Generation Is Entering Workflow
Contemporary scribes extend into multiple workflow functions. Problem list suggestions — identifying clinical problems and proposing additions to the permanent record. SNOMED CT coding — mapping language to structured codes with register, QOF, and prescribing implications. Referral letter drafting — generating content with clinical prioritisation (routine vs urgent vs 2-week wait). Investigation suggestions — proposing tests based on the clinical discussion. Follow-up task creation — generating tasks for results review, callbacks, and reassessment. Patient messaging — generating post-consultation communications. Safety-netting — producing red-flag advice and re-presentation criteria.
Each extension adds genuine value. Each also moves the scribe further from documentation toward clinical decision support — and closer to a different regulatory and accountability framework.
The Boundary Problem
When does a documentation tool become a clinical decision support tool? The answer matters for regulatory classification, clinical governance, and professional accountability.
Is suggesting a SNOMED code decision support? A code is a clinical assertion. Suggesting "asthma" rather than "possible asthma" is a diagnostic judgment. If accepted without independent evaluation, the AI has influenced the permanent record.
Is drafting a referral decision support? Referral letters include clinical prioritisation, information selection, and implicit recommendations about urgency. These are clinical judgments embedded in a document that the receiving clinician will act upon.
Is suggesting an investigation decision support? "Request HbA1c and fasting glucose" is an investigation recommendation. If ordered because the scribe suggested it, the AI has contributed to a clinical decision with real consequences — resource use, patient anxiety, incidental findings, and follow-up obligations.
Is generating safety-netting advice clinical advice? "Return if headache becomes sudden-onset or you develop neck stiffness" implies a specific differential (subarachnoid haemorrhage) and a re-presentation threshold. If the safety-netting is incomplete — the AI generates generic "return if symptoms worsen" when the presentation warranted specific red-flag advice about neck stiffness, photophobia, and thunderclap onset — the clinical risk is created by the AI's omission. The patient receives inadequate safety-netting that looks adequate in the record.
Is drafting a patient message clinical communication? A post-consultation message to the patient that summarises the management plan, explains what to watch for, and provides return criteria is clinical communication — not merely administrative correspondence. If the AI generates a message that omits a key caveat discussed during the consultation, the patient receives incomplete information from what appears to be a clinician-authored communication.
Is this a medical device function? Under MHRA guidance, software providing information intended to inform clinical decisions about individual patients may meet the definition of a medical device. As scribes evolve to include investigation recommendations, treatment suggestions, and referral prioritisation, the regulatory classification question becomes unavoidable. A tool classified as documentation software faces different regulatory requirements than a tool classified as clinical decision support — even if both tools generate the same types of outputs.
The Practical Impact on Different Clinical Settings
The boundary problem manifests differently across settings. In general practice, where the GP makes independent decisions about investigation, prescribing, referral, and safety-netting — all within a single consultation — AI workflow suggestions have immediate clinical impact with no intervening review layer. In hospital settings, workflow suggestions may be reviewed by a senior clinician before implementation, providing an additional verification layer. In urgent care and emergency departments, time pressure may reduce the clinician's capacity to critically evaluate AI suggestions, increasing the risk of uncritical acceptance. In community and domiciliary settings, clinicians working alone without immediate colleague access may be more likely to rely on AI workflow support — making the quality and safety of that support more consequential.
Why Clinicians Need to Understand the Boundary
The accountability model differs between documentation and decision support. A documentation error is a verification failure — the clinician failed to check the draft. A decision support error raises questions about the tool's clinical validation, the manufacturer's safety case, and the regulatory classification. Clinicians should ask: is this tool helping me write the note, or helping me make the decision? If both — which outputs are documentation and which are recommendations? Am I verifying both with appropriate rigour?
The Missing Layer: Source-Grounded Verification
A note can be well written but clinically incomplete. A referral can be polished but miss guideline-required information. A coding suggestion can be plausible but diagnostically wrong. The missing layer is source-grounded clinical verification — checking outputs against authoritative guidance. "What does NICE actually recommend?" "Which red flags should be included?" "Is this the right referral pathway?" "Does this code reflect the actual diagnostic certainty?"
This verification is not something the scribe provides — it requires a different tool that retrieves, cites, and supports clinical reasoning.
A Safer Workflow Model
- Consultation occurs.
- AI scribe drafts documentation — note, codes, letters.
- Clinician reviews — accuracy, completeness, uncertainty preservation.
- Clinician checks uncertain questions against trusted guidance.
- Clinician verifies safety-netting — specific, time-bound, guideline-informed.
- Clinician confirms codes — matching what was assessed.
- Note saved — clinician-verified, clinician-owned.
- Learning captured as CPD where appropriate.
The scribe serves steps 2-3. iatroX serves steps 4, 5, and 8 — the verification, safety-netting, and learning layers.
Use iatroX after the note, during the question, or before the referral →