After AI takes a history, the doctor should verify the consequential information, resolve what remains unknown and establish whether the account still fits the patient. The choice is not between repeating every question and accepting the summary unchanged. AMIE's supervised pre-visit study makes this handover a practical clinical question rather than an abstract debate about replacing doctors.
In the Lancet abstract published on 8 October 2026, transcripts and summaries were shared before the patient's usual appointment. The review process proposed below is an editorial workflow for thinking about such material, not a process proven superior by that study.
Start by identifying what kind of statement you are reading
A polished summary can contain patient-reported information, an AI interpretation, a fact from a record and an unresolved question. Those statements should not be treated as though they have the same evidential status.
"The patient reports no previous episode" is different from "no previous episode appears in the record". Both differ from a summary that simply omits previous episodes. A clinician needs to know which statement is being made before deciding whether it is sufficient for the next decision.
A useful handover would therefore separate the source of information from the conclusion drawn from it. It should preserve uncertainty, identify important unanswered questions and make it possible to inspect the relevant part of the original conversation. These are proposed design requirements, not claims that every AMIE summary already meets them.
A fictional summary that looks complete
Consider this constructed teaching example: "Fatigue is consistent with poor sleep. No relevant associated symptoms. The patient wants reassurance."
Now examine the fictional underlying conversation. The patient described poor sleep but also said the fatigue was changing. The chatbot did not explore what had changed. The patient never explicitly denied associated symptoms, and they asked whether an existing medicine might be contributing. The summary has converted missing information into a negative finding and narrowed the patient's goal.
The next clinician task is not to select an investigation from this paragraph. It is to recognise that the evidence does not yet justify the tidy interpretation. The learner reviewing the case should identify which statements require clarification and explain why each could change the assessment.
This is not an AMIE incident or a report of a current iatroX output. It is a deliberately constructed exercise in distinguishing a plausible narrative from a sufficiently established account.
Not mentioned is not the same as absent
An explicit negative answer can be useful when the question was understood and was appropriate to the clinical decision. An omission may instead mean that the topic was not asked about, was not answered or was lost during summarisation.
Treating these states as equivalent can change the reasoning. A clinician may decide not to ask a relevant question because the summary appears to have answered it already. The issue is therefore not merely documentation style; it concerns what information remains to be obtained.
The practical response is selective verification. Check the negatives and assumptions that carry the plan, rather than demanding that every inconsequential detail be repeated. When the underlying conversation is unavailable or ambiguous, the limitation should remain visible.
For the broader cognitive issue, iatroX's existing article on pre-visit summaries and anchoring examines how an initial account could shape attention. The present task is more specific: deciding what to verify at the next encounter.
A proposed handover review
| Review step | Question to ask | Useful output |
|---|---|---|
| Re-establish the patient's purpose | What matters most to the patient today? | A current priority rather than an assumed agenda |
| Check consequential facts | Which statement would change the plan if wrong? | Targeted clarification rather than wholesale repetition |
| Separate absence from omission | Was this explicitly asked and answered? | A visible unresolved question where needed |
| Look for contradictions | What does not fit the prepared interpretation? | A reason to broaden or revise the assessment |
| Decide what requires direct assessment | Is examination or another source necessary? | A boundary on what text alone can establish |
| Make the next decision explicit | What happens now, and who is responsible? | A current plan with ownership and reassessment |
This is an original review framework. It should be tested for usability and clinical effect before anyone presents it as a validated supervision protocol.
Keep the patient in the handover
The patient is not simply the source of data that has now been handed to the clinician. They can correct the account, describe a change and explain what they understood from the AI conversation. Beginning with their current priorities can expose a discrepancy that a technically accurate transcript would not resolve by itself.
For example, a fictional patient might say that an AI-mentioned possibility is now their main concern, even though it was a minor part of the generated list. The clinician needs to address that interpretation while assessing the underlying problem. Ignoring it because the differential is technically defensible would leave part of the consultation unfinished.
The GMC's decision-making and consent guidance, checked on 10 October 2026, supports dialogue that takes the individual's concerns and understanding seriously. It does not require a particular AI review template, but it provides the professional purpose for the conversation.
Do not confuse verification with reconstructing everything
A well-designed summary could reduce repetition and expose information the clinician might otherwise miss. The appropriate response is not a blanket rule that every AI-collected history must be retaken word for word.
Instead, match the verification to the consequences. A detail that determines whether the current interpretation is plausible deserves attention. A clearly documented, patient-confirmed background detail may not need to consume the same amount of consultation time. The clinician should be able to explain that distinction.
AMIE's March 2026 methods describe different information access for the AI and the treating clinician. The preprint therefore reinforces the need to extend the account where examination, records or later information matter. Verification is part of moving the case forward, not just checking whether the chatbot transcribed correctly.
How would we know whether this role improves care?
A useful comparison would examine the clinician-AI partnership rather than only the summary. Does selective review catch important omissions? Does it reduce unnecessary duplicate questioning? Does the patient feel able to correct the account? Are subsequent decisions more appropriate, or simply faster?
The evaluation should also test an apparently excellent summary that contains one consequential assumption. If reviewers can only recognise obvious nonsense, the workflow may look reassuring without demonstrating effective challenge. Conversely, a good review process should let clinicians use accurate material efficiently, rather than rewarding indiscriminate rejection.
These are proposed research and teaching questions. iatroX's reasoning tools can provide an educational setting for fictional cases and source checking, but no AMIE result establishes that any particular learning platform makes supervision safe.
Frequently asked questions
Should the doctor repeat the entire AI history?
Not automatically. The proposed approach is to verify consequential information, clarify gaps and obtain whatever direct assessment is required for the current decision.
What is the most important difference between an omission and a negative finding?
A negative finding should reflect information actually established; an omission may mean the issue was never explored. Treating an unknown as absent can distort the assessment.
Has this handover framework been validated in the AMIE study?
No. It is an editorial proposal for clinical review and teaching, not a reported intervention or validated outcome of the study.
Practise identifying the next question in an incomplete case →