Two AI categories now sit in the consultation room, and confusing their jobs is becoming the characteristic new error of digitised practice. A scribe records and structures what was said; an evidence tool helps answer what should be considered; and the danger zone is the join, where a transcript's contents flow toward a prescribing decision without the checking that either category alone would have received. This article separates the four layers, walks the risk chain that runs through them, and sets out the workflow that keeps an ANP's consultation fast without letting the machinery decide anything.
The four layers, and what each is accountable for
Consultation capture: the ambient recording and transcription layer, where the accountable property is fidelity, did it hear what was said, and where the regulatory position is settled and clarifying: transcription-and-summary tools for clinician review are not medical devices precisely because the clinician's review is the control. Note generation: the structuring layer, where the accountable property is faithfulness to the encounter plus the clinician's correction, because the generated note is a draft until reviewed and a record after signature, and the signature is the ownership transfer. Clinical evidence retrieval: the knowledge layer, where the accountable properties are the ones this cluster measures everything by, grounding, jurisdiction, dates, inspectable sources, and where the question should be asked deliberately rather than inherited from the transcript's phrasing. And the prescribing or ordering workflow: the action layer, governed by the clinical system's own checks and the prescriber's authorisation, which no upstream layer's output should reach unexamined. Products increasingly span layers, Heidi's evidence function beside its documentation being the visible example, and spanning is fine; the accountability at each layer is what must not blur, per the fuller comparison at /blog/iatrox-vs-heidi-evidence-anp.
The risk chain, link by link
The failure this architecture exists to prevent is propagation, and it is worth seeing whole: a medicine name mis-transcribed in capture; a dose or frequency misheard and rendered confidently in the note; an allergy mentioned but dropped by the summariser; the generated note then treated as confirmed fact rather than draft; and the error inheriting forward into the prescription, the referral letter and the patient's instructions, each step laundering the mistake with more formatting. No single link is exotic, every link is documented in incident literature's pre-AI equivalents, and AI changes the chain's speed more than its nature. The defence is also link-by-link: medication identity and allergies verified by the clinician against the patient and record, not the transcript; the note reviewed as a draft with the high-risk fields, medicines, doses, allergies, numbers, checked deliberately rather than skimmed; and nothing in the action layer initiated from generated text alone.
The safe workflow, and the honest configuration
Five steps that cost seconds and break the chain. The clinician confirms history and medication identity directly, the human layer no tool replaces. The evidence question, where one exists, is asked separately and cleanly, stated by the clinician rather than inherited from transcript phrasing, because a well-formed question is half the answer's safety, and it goes to the layer built for it, source-grounded, jurisdiction-correct, dates visible. Supporting sources are inspected per the 90-second screen, /blog/verify-ai-prescribing-answer-90-seconds. The final plan is documented and authorised independently, the prescriber's own act on the prescriber's own verification. And the patient is informed of the tools in use where local policy requires, the transparency layer that keeps the room honest. The configuration this implies is the one the comparison article reaches: a locally approved scribe for capture and documentation, a dedicated evidence layer for knowledge and verification, iatroX's position in the pair, and the join between them staffed, permanently, by the clinician, because the workflow's safety was never in either product; it is in the checking that the division of labour makes cheap enough to actually happen.
Frequently asked questions
Doesn't reviewing everything erase the scribe's time savings?
No, because review is targeted: the high-risk fields get deliberate checking, the narrative gets a read, and the net saving survives comfortably; the alternative, unreviewed notes, spends the saved time later with interest.
Should evidence functions inside scribes be avoided?
Used with the same standards, not avoided: the question deliberately formed, the sources inspected, the automation-bias pressure of in-flow convenience consciously resisted; the architecture changes where the discipline is applied, not whether.
What about consent for recording itself?
Local policy governs and transparency is the safe default: patients told a documentation tool is in use, objections respected, and the practice's information-governance route consulted for the specifics, which vary by organisation.
Who is accountable if a scribe error reaches a prescription?
The signing clinician, which is the architecture's point: the review-before-signature step is where accountability is exercised, and incident reporting should still capture the tool's contribution so the system learns what the signature caught.