Training clinicians to supervise AI should be a shared responsibility of product vendors, healthcare organisations and educators. Vendors must explain actual product behaviour, organisations must define local review and escalation, and educators can create practice that tests judgement. Heidi and Tandem's September 2026 expansion plans make that division of responsibility more important, not less necessary.
This is an educational proposal, not a report of an established joint training programme. The article is published by iatroX and includes iatroX's existing learning tools while distinguishing them from a new, dedicated curriculum for supervising other suppliers' agents.
Supervision needs a defined decision
Heidi's September 2026 announcement describes supervised work around clinical care, while Tandem's 14 September 2026 announcement describes expansion towards broader clinic operations. Those are product directions, not evidence that clinicians have already been trained to supervise every proposed function. Heidi's announcement and Tandem's roadmap provide the starting point.
The first educational task is to identify what the reviewer is deciding. Reviewing whether a note faithfully reflects a conversation is different from deciding whether a proposed follow-up should be sent. Checking a cited recommendation is different from confirming that a task reached the right destination. A generic instruction to "check the AI" leaves too much unspecified.
For each activity, training should explain the information available to the system, the information it may lack, the permitted action and the point at which the clinician can reject or amend it. That makes supervision teachable and assessable.
Four fictional situations worth practising
An incomplete note can appear coherent while omitting the patient's uncertainty about a proposed plan. A useful exercise asks the learner to compare the draft with a synthetic transcript and identify where the meaning changed. The task is not to reward elegant rewriting; it is to preserve the clinically relevant distinction.
A second scenario presents a recommendation generated from incomplete background information. The learner should explain which missing facts could change the decision and why the proposed action should pause. The exercise tests whether the learner recognises uncertainty rather than rewarding confident acceptance or automatic rejection.
A third scenario includes a credible citation that does not support the particular sentence attached to it, or guidance drawn from a different jurisdiction. The learner must inspect applicability and support, not merely notice that a reference is present.
A final scenario proposes a follow-up communication that should not proceed because the intended recipient or approval state is unresolved. The learner must decide where the work stops, who owns the exception and how the unresolved issue is made visible. These are invented training scenarios without patient-identifiable information, not reports of errors by Heidi or Tandem.
Compare outputs before revealing an explanation
An educator could present two plausible drafts and ask the learner to choose, justify the choice and identify weaknesses in both. A later discussion could reveal why the more fluent output was not necessarily the better one. The educational aim is to make the reasoning visible before feedback is supplied.
Another exercise could keep the proposed action constant while changing one piece of context. The learner then explains whether the action remains appropriate. This tests whether they are applying the underlying distinction rather than memorising an example.
A third activity could ask the learner to disagree with a draft constructively: identify the unsupported statement, name the information needed and specify the appropriate review route. Rejection alone is not enough if it leaves colleagues uncertain about what should happen next.
Confidence is not the same as competence
A satisfaction survey can tell an educator whether learners found training understandable. It cannot establish that they can detect a missing qualifier, identify an irrelevant source or stop a task at the right point. The assessment should require those behaviours directly.
A proposed assessment would record the learner's initial judgement, explanation and response to feedback. It would distinguish a justified acceptance from an uncritical one, and a justified refusal from blanket distrust. A later unfamiliar scenario could test whether the learner can apply the same reasoning beyond the training example.
No such assessment was run for this article, and no pass-rate or competence result is claimed. Results should be published only after an actual study or documented training evaluation, with the participants, cases, scoring criteria and limitations stated.
What vendors and employers should supply
The vendor should explain the current product's inputs, limitations, review controls and behaviour when a task cannot be completed. Training needs to change when those behaviours change. A generic introductory course cannot substitute for knowing what a specific release does in the local configuration.
The employer should clarify approval, documentation and escalation responsibilities. Learners need to know who handles an unresolved transfer, how a suspected error is reported and what fallback process remains available. These are proposed operational requirements, not a legal opinion about a particular deployment.
Educators can then build scenarios around that verified workflow. The resulting course is more useful than abstract "AI awareness" because the learner practises decisions they are actually expected to make.
Where iatroX's existing tools could contribute
Per iatroX product information, September 2026, its Socratic Tutor begins with an attempted question, asks targeted follow-ups and aims to identify the learner's misconception. Adaptive question banks and spaced repetition support structured revisiting of material. Those features could support transferable reasoning practice, but their presence is not evidence that a clinician can supervise a particular vendor's agent. iatroX Tutor and the learning entry point describe the existing tools.
At launch in September 2026, iatroX Simulations comprised eighteen examination-specific tracks and 1,114 clinician-reviewed cases. Per iatroX product information, September 2026, the service includes voice and text, practice mode with pause-and-coach, uninterrupted exam mode, full mock circuits, transcript-linked feedback by domain, Tutor-led remediation, a prescribed next case and shared progress across web and native apps. One complete simulation is free. iatroX Simulations is the existing product area.
Those are examination-practice design features. Clinician review does not mean endorsement by an examining body. The simulations are not presented here as clinically validated supervision training, pass-rate predictors or replacements for bedside and procedural practice. A dedicated Heidi-and-Tandem supervision curriculum would be a new educational proposition requiring product-specific content and evaluation.
Turn practice into a personally reviewed record
Per iatroX product information, September 2026, its CPD workflow supports profession and practice-context selection, personal review of a draft record, PDF export and direct FourteenFish export for linked accounts. Completed evidence remains available after the subscription ends. CPD and simulations are included with the paid learning subscription, not separate add-ons. iatroX CPD describes the record pathway.
A useful record would explain the error pattern encountered, the reasoning corrected and what the learner will do differently. It should not simply state that an AI session was completed. Nor should an exported record be described as formally accredited CME without the relevant accreditation basis.
The practical educational conclusion
For product-specific supervision, start with vendor behaviour and local responsibilities. For transferable clinical reasoning, use structured practice that requires an explanation before feedback. For employers, assess performance on relevant scenarios rather than relying on confidence alone. More capable software creates a need for clearer judgement at the points where a human remains responsible for review.
Frequently asked questions
Does iatroX already provide a dedicated Heidi or Tandem agent-supervision curriculum?
No such curriculum is claimed here. The proposal would be a new educational product, distinct from iatroX's existing examination-focused tutoring, questions and simulations.
Can a confidence survey establish safe AI supervision?
No: confidence and satisfaction do not directly test whether a clinician can identify omissions, assess sources or stop an inappropriate action. Scenario-based assessment should test those decisions.
Are clinician-reviewed simulations the same as accredited training?
No: clinician review, examining-body endorsement and formal educational accreditation are separate claims. A learning record should describe the activity honestly without implying an accreditation it does not have.