Doctors supervising healthcare agents need more than familiarity with AI terminology. They need to identify missing information, challenge a proposed action, recognise a poor hand-back and explain changes to a patient. These are tasks that can be practised. They should not be assumed merely because a clinician knows how to use a chat interface.
This article proposes an educational approach as of 9 October 2026, not an established mandatory competency standard. AgentClinic, published in April 2026, and the July 2026 PatientAgentBench preprint provide a rationale for evaluating interactive work rather than relying only on static answers. Neither establishes that the teaching design below has been validated.
Define the supervision task before choosing the teaching tool
Supervision can involve different decisions. A clinician might check whether the history is sufficient, assess a proposed plan, authorise an external action or take over after an interruption. Each requires different information and should be assessed separately.
A useful teaching objective is observable: the learner identifies the unsupported assumption and states what would resolve it. A less useful objective is simply to increase confidence with AI. Confidence can improve without the learner becoming better at detecting a consequential problem.
The cases below are original fictional exercises. They are not reports of incidents, reproductions of patient encounters or claims that these exact scenarios are already available in the iatroX catalogue. Their purpose is to make supervision concrete enough to teach and evaluate.
Case one: the history that looks complete but is not
A virtual intake summary describes a patient with a recurring headache and concludes that there is no relevant associated history. In the underlying conversation, however, the patient did not answer one important question. The summary has converted missing information into an apparent negative.
Ask the learner to mark what is known, what was explicitly denied and what was never established. The task is not to guess a diagnosis from an incomplete record. It is to explain why the proposed plan cannot yet be assessed and to request the specific information that would change the decision.
The facilitator should reward a targeted correction rather than a demand to repeat the entire encounter. The important learning point is that faithful-looking prose can conceal a change in evidential status. An unanswered question should remain unresolved until appropriately addressed.
Case two: a plausible plan that does not fit this patient
A fictional patient receives a generic proposal to repeat an earlier management approach. The record also contains a previous reason why that approach was stopped. The generated plan is coherent and supported by general information, but it does not address the patient's documented experience.
Ask the learner to explain the mismatch without simply declaring the AI wrong. Which part of the plan is reasonable in general? Which patient-specific fact changes its applicability? What needs clarification before a revised decision can be made?
A strong response separates source support from suitability. The learner should also explain the change in language a patient can understand. The educational outcome is not a longer critique; it is a defensible, proportionate reassessment of the proposed action.
Case three: the action marked complete without confirmation
A practice agent says a referral has been sent. The available trace shows that a document was prepared and a submission attempted, but the destination returned an ambiguous response. There is no reliable acknowledgement of receipt.
Ask the learner to distinguish drafted, authorised, submitted and accepted. They should identify what must be checked before repeating the action, because an uncertain response can coexist with a successful external submission. Blindly retrying could create duplication.
The correct supervision behaviour is to establish the actual state and communicate uncertainty accurately. The scenario draws on a failure mode discussed in the July 2026 PatientAgentBench research, but it is not a measured result for a named commercial product. The teaching case tests recovery, not merely the recognition of a wrong sentence.
Case four: two sources describe different plans
A patient-facing summary recommends continuing the previous follow-up arrangement. A later clinician-authored record changes that plan after new information. Both documents are available, but the agent has treated them as equally current and produced a blended instruction.
Ask the learner to identify the source, date and authority of each statement. The answer should not be whichever document sounds more confident or appears last in the interface. The learner must establish which decision is current and whether any completed or pending action now needs correction.
The final task is to explain the current plan to the fictional patient and identify who owns the unresolved work. This tests reconciliation and communication, not only factual recall. It also exposes whether the learner notices that correcting a note may leave an earlier external action unchanged.
Assess the explanation, not the approval speed
For each case, ask what the learner accepted, challenged or changed and why. Record the evidence they used, the missing information they identified and whether their proposed next step fits the authorised role. Clicking approve quickly is not a measure of sound supervision.
Assessment should also recognise proportionate action. Escalating every case or demanding every possible investigation can avoid commitment without demonstrating judgement. A good learner should distinguish what is necessary now from what would be interesting but irrelevant to the decision.
The GMC's prescribing guidance, checked on 9 October 2026, provides a professional anchor for sufficient information before treatment. The delegation and referral guidance provides another for clear transfer of work. Neither turns this proposed teaching exercise into a formal qualification.
Turn an error into a new attempt
A useful remediation sequence begins with the underlying gap. Was the learner missing clinical knowledge, misreading the evidence, overlooking the action status or failing to identify ownership? The next activity should address that gap rather than repeat an identical case until the expected answer is memorised.
After a focused explanation, use a contrasting scenario. A learner who over-escalated should encounter a case where completion is justified. A learner who accepted an unsupported negative should encounter a case where the negative was properly established. Later reassessment without the same assistance can explore whether the distinction transfers.
This is an educational design to evaluate. It is not evidence that any particular app prevents deskilling or that improved simulation performance guarantees safe supervision in practice. Human observation and feedback remain important, especially for communication and professional judgement.
How iatroX can contribute without overstating the claim
Under iatroX's September 2026 specification, Socratic Tutor opens on an attempted question and uses targeted follow-ups to explore the learner's misconception. Question banks add adaptive sequencing and spaced repetition. These functions can support deliberate work on the reasoning identified during a supervision exercise.
At its September 2026 launch, iatroX Simulations included eighteen examination-specific tracks and 1,114 clinician-reviewed cases, with voice and text, coached practice, uninterrupted exam mode and transcript-linked feedback. That is a description of educational content and design, not examination-body endorsement, clinical validation or permission to supervise another AI product.
The September 2026 subscription is £99 paid upfront for a year, equivalent to £8.25 a month billed annually, or £29 monthly. It includes question banks, Socratic Tutor, the study planner, Simulations and CPD tools together; Simulations and CPD are not separate add-ons. Three monthly payments total £87 and four total £116, so annual payment is cheaper from the fourth month.
Ask-iatroX and free question access remain genuinely free without trial expiry or a verification gate, and one complete simulation is free under that specification. The value proposition is several learning methods serving one relevant professional goal, not buying access to unrelated examinations.
Keep learning evidence separate from certification
A clinician can document a genuine learning activity, what changed in their reasoning and what they will practise next. Under iatroX's September 2026 CPD specification, the learner reviews and personalises the draft record, with PDF export and direct FourteenFish export for linked accounts; completed evidence remains accessible after the subscription ends.
That record is evidence of learning, not automatically accredited CME or proof of competence to supervise a particular deployment. The most useful outcome is a clinician who can explain when to accept an action, when to challenge it and how to take over without losing the patient's actual problem.
Frequently asked questions
Is healthcare-agent supervision already a single mandatory qualification?
This article does not establish such a standard. It proposes specific educational tasks that organisations can evaluate alongside their professional and deployment responsibilities.
Does iatroX certify doctors to supervise healthcare agents?
No such certification is claimed. The September 2026 product specification describes reference, learning, simulation and CPD tools, not approval to supervise another company's system.
What should a supervision exercise measure?
It should assess whether the clinician identifies missing information, evaluates the action and explains an appropriate response. Approval speed or confidence alone is insufficient.