An AI scribe records information elicited during an encounter. An AI intake system helps decide which information to obtain. That change alters the evidence required: a highly accurate transcript cannot establish that the right questions were asked. AMIE's supervised real-patient feasibility study brings this distinction into a concrete clinical workflow. The Lancet abstract published on 8 October 2026 describes conversations before ordinary primary-care appointments.
The implication is not that documentation research has become irrelevant. It is that evidence should follow the function being delegated. A product that expands from recording a consultation into conducting part of it needs an evaluation of that additional work.
A perfect record of an incomplete conversation
Consider a fictional encounter in which a patient describes a new symptom. The discussion covers its duration and effect on daily life, but never establishes whether a recent change in medicines could be relevant. A scribe could reproduce the conversation without error and still leave the resulting record incomplete for the clinical decision.
That would not automatically be a transcription failure. It would be a limitation of the information-gathering process. If an intake system was responsible for asking the questions, however, its selection and sequencing of questions would become part of the assessment.
Now imagine that the intake system asks about medicines but accepts an ambiguous answer without clarification. The transcript may remain accurate while the summary overstates what was established. This creates another distinct task: preserving the status of information between conversation and handover.
These are constructed examples, not documented failures in AMIE, TORTUS or any other named product. Their purpose is to distinguish the claims that an evaluation should test.
What AMIE adds to the evidence discussion
In the studied pathway, AMIE conversed with patients under real-time physician supervision, and a transcript and summary were shared with the clinician conducting the subsequent appointment. Google's account updated on 8 October 2026 makes the human involvement and sequence explicit.
The system was therefore doing more than recording an encounter led entirely by a clinician. Its questions helped determine the information available in the generated account. The study examined the conversational process and user experience, rather than assuming that fluency or transcription quality was sufficient.
The result remains specific to that supervised arrangement. It does not validate every patient-facing agent, and it does not establish that a product with strong scribing results can inherit the same performance on intake. The same principle applies in reverse: useful intake evidence does not prove accurate prescribing or reliable execution of a follow-up action.
A task-by-task evidence matrix
The following is an editorial vendor-assessment framework. It should be populated with evidence for the exact feature and deployment, not a single company-wide badge.
| Claimed capability | Evidence to request | What a narrower evaluation would miss |
|---|---|---|
| Transcription and documentation | Fidelity, omissions, attribution and preservation of uncertainty | Whether the encounter elicited the information needed |
| Patient intake | Relevant questions, clarification, out-of-scope recognition and comprehension | Whether a polished summary rests on an incomplete history |
| Clinical recommendation | Appropriateness given patient context and supporting evidence | Whether a plausible explanation justifies this decision |
| Workflow execution | Authorisation, correct action, acknowledgement and recovery | Whether the action was clinically appropriate in the first place |
| Clinical reference | Source relevance, support for claims and applicability | Whether information was acted on appropriately |
| Professional learning | Feedback quality and subsequent demonstrated understanding | Whether an improved practice result transfers to clinical work |
This comparison is published by iatroX, which includes its own reference and learning functions in the last two categories. They are not substitutes for a patient-treatment service, and the AMIE study does not validate their effectiveness. Product descriptions and evaluation questions in this article are current to 10 October 2026.
Test information gathering, not just the final answer
A diagnostic label at the end of a conversation does not reveal whether the route was robust. The system might have obtained the relevant information, guessed successfully from limited details or relied on a misleading assumption that happened not to change the final label.
A useful evaluation would therefore inspect the process. Did the system clarify an ambiguous answer? Did it recognise a conflict between earlier and later information? Did it distinguish an explicit negative from an unanswered question? Did it stop collecting detail when another form of assessment became necessary?
These are proposed questions, not claims that the AMIE study measured every one. They help explain why a static examination score cannot stand in for an evaluation of interactive intake.
The handover is another clinical task
Information must remain usable when it moves from the patient conversation to the receiving clinician. The relevant handover should preserve the patient's main concern, the source of important facts, unresolved uncertainty and any change that requires reassessment.
The receiving professional also needs to know what was not available to the AI. A conversation-only summary should not imply that a longitudinal record has been checked. If a clinician assumes that it has, a documentation feature can acquire influence beyond what its description appears to promise.
iatroX's related article on evidence for TORTUS Health beyond scribing asks the same portability question for a named care proposition. The AMIE lesson is broader: evaluate the information-gathering component, the human-AI interaction and the complete service separately.
How buyers should handle an expanding product claim
Ask the supplier to identify the exact task supported by each cited study. Record the population, system version, available information, human review and measured outcome. Then compare those fields with the proposed local use.
A supplier may have excellent evidence for one feature and preliminary evidence for another. That is more useful than either treating the whole product as clinically validated or dismissing all its functions because one is still developing. The deployment decision should be proportionate to the evidence and the consequences of the task.
The DECIDE-AI reporting framework is relevant to early live evaluation because it addresses human factors and clinical workflow. It is not a certificate, and compliance with a reporting framework is not the same as proving that a service improves outcomes.
What counts as progress?
For a documentation buyer, strong evidence of accurate, useful notes may be directly relevant. For an intake buyer, the additional questions concern completeness, patient interaction and the next clinician's decisions. For a clinical educator, the distinction suggests teaching learners to examine how information was obtained, not only whether an answer sounds plausible.
The aim is not to draw permanent boundaries around companies. A single product may legitimately span several functions. Each expansion should make the evidence map more specific, rather than stretch a general claim until it covers tasks that were never evaluated.
AMIE changes the conversation by providing a real example of supervised AI participation before the clinical encounter. The next step is to evaluate what that participation changes across the pathway, not merely celebrate that the conversation can be generated.
Frequently asked questions
Why is intake different from AI scribing?
Intake helps determine which questions are asked and what information becomes available. A scribe can accurately record an encounter without establishing that the encounter collected everything needed.
Can a product's scribing study validate its clinical agent?
Not automatically. Evidence needs to match the added functions, such as history gathering, recommendation, execution and escalation.
What should buyers ask when a vendor says clinically validated?
Ask which exact feature, system version, population, workflow and outcome were evaluated. A broad label is not a substitute for those details.