The best consultation automation removes avoidable handling while preserving the understanding needed to make and explain a clinical decision. Information collection, documentation and task preparation may be assisted. Judgement about uncertain information, the patient's priorities and the appropriateness of a plan still needs accountable human involvement within the intended service.
This is a proposed practice framework, not a universal regulatory classification or a claim that every consultation should use AI. The professional guidance and product examples were checked on 15 October 2026.
Break the consultation into work, not job titles
A consultation includes listening, clarifying information, interpreting findings, considering alternatives, agreeing a plan, documenting it and arranging what follows. Some of these activities are mechanical; others depend on context that may be incomplete or difficult to express.
The right question is not whether AI can "do the consultation". It is which task is being delegated, what information the task requires and what a person must still understand before relying on the output.
A system can influence care without making the final decision. Selecting which historical details appear in a summary can shape the clinician's attention. Drafting a patient explanation can change how uncertainty is communicated. Those functions deserve assessment even when labelled administrative.
A fictional encounter begins with a mismatch
A fictional patient attends to discuss persistent discomfort. The pre-visit summary emphasises a familiar diagnosis from the record. When the consultation starts, the patient explains that their main concern is a new limitation that was not captured by the form.
The clinician should not treat the task as confirming the prepared story. The summary may still be useful, but the patient's current account changes the question. Automation has helped organise information without establishing which information matters most now.
This example is deliberately about a change in meaning rather than an invented fact. An AI can accurately summarise the material it received and still provide an inadequate starting point if the interaction failed to elicit the patient's actual priority.
Match assistance to the consequence
| Consultation task | Potentially useful assistance | Human work that must remain visible |
|---|---|---|
| Information gathering | Organising a history and identifying unanswered fields | Clarifying ambiguity, changing the question and deciding what further assessment is needed |
| Clinical reasoning | Surfacing relevant evidence or alternatives | Judging applicability, contradictions and uncertainty |
| Shared decision-making | Preparing accessible explanations | Understanding preferences and agreeing a plan with the patient |
| Documentation | Drafting an account for review | Confirming what was actually said, observed and decided |
| Follow-through | Preparing authorised tasks | Approval where required, exception handling and ownership of unresolved work |
This is an editorial allocation framework. The suitable level of assistance depends on the product, patient, setting and relevant evidence, not on the table alone.
The missing information problem
A text or audio system cannot safely infer every finding that would require examination or further questioning. It should distinguish information that is absent from the record from information explicitly assessed and found absent.
In the fictional encounter, the clinician may need to examine the patient, clarify chronology or consult another professional. An answer generated from the original summary cannot remove those needs simply by being detailed.
The GMC's guidance on innovative technologies makes professional judgement and the limits of competence central. The practical implication is to ask what information could materially change the plan before treating a prepared recommendation as ready to approve.
Review should test meaning, not polish
An AI-generated note can be easy to read while altering the status of a statement. A possibility can become a confirmed diagnosis. Advice that would have been sensible can appear as advice actually given. A conditional plan can look completed.
A meaningful review checks attribution, chronology, uncertainty and the agreed next steps. It does not merely improve grammar or remove repetitive text. The reviewer also needs enough time and access to the original information to perform that work.
The MHRA's July 2026 ambient-voice clarification distinguishes documentation-only functions from other clinical or autonomous functions in Great Britain. That regulatory distinction does not mean documentation errors are clinically irrelevant or that every function under one brand has the same status.
Avoid both blind approval and unnecessary duplication
The answer is not to repeat every automated step manually. If a tool removes no work because the clinician reconstructs everything from scratch, its value may be limited. Equally, clicking approve without assessing consequential content is not meaningful supervision.
The appropriate design should expose the information needed for the particular review. A draft letter should show the relevant source material and destination. A clinical suggestion should make important assumptions and uncertainty visible. A completed task should have confirmation from the receiving system.
This is a product and workflow problem as much as an individual behaviour problem. A warning that "the clinician is responsible" does not provide the information, authority or time required to exercise that responsibility.
Patient transparency must be specific
The patient should understand what the tool does in this encounter. Recording and drafting a note is different from asking questions, suggesting a plan or sending follow-up communication. A vague statement that the practice "uses AI" does not explain those distinctions.
The clinician should also be able to explain alternatives and limitations without promising perfect security or accuracy. The relevant professional discussion is supported by GMC guidance, while the exact information-governance and consent arrangements depend on the service and jurisdiction.
In the fictional consultation, a patient who does not want a particular recording tool should not have to argue about whether AI is good in general. The practice needs a usable process for the actual preference and an appropriate alternative pathway.
Practise the judgement that automation does not remove
A useful training exercise presents the learner with a plausible but incomplete summary and asks what remains unresolved. A second exercise asks them to explain why they changed a proposed plan. A third tests recovery when an action is reported complete but its status is uncertain.
These are proposed educational activities, not validated prevention of automation bias. They should be conducted with fictional or appropriately de-identified material and appropriate supervision.
As described for September 2026, iatroX Tutor uses follow-up questions around an attempted answer, and its simulations provide another practice format. Those tools can support reasoning practice but do not certify a clinician to supervise a particular AI product or replace bedside assessment.
The standard is better care with understandable responsibility
A useful consultation assistant should leave the clinician better able to attend to the patient, not merely produce a more polished record. The service should evaluate the whole encounter, including omissions, correction work, patient understanding and subsequent actions.
Automating more is not a sufficient objective. The appropriate target is a consultation in which avoidable handling is reduced and the people involved still understand the problem, the plan and what to do when the situation changes.
Frequently asked questions
Should every part of a consultation be automated if technically possible?
No: technical capability does not establish clinical suitability. Delegation should follow the task, evidence, patient context and consequences of error.
Is clinician approval enough to make an AI output safe?
Not by itself: the reviewer needs relevant information, time and authority to challenge the output. Approval records do not measure understanding.
Can AI simulation replace clinical experience?
No: it can supplement practice in reasoning and communication, but it does not replace supervised bedside or procedural learning.