Doctors will need to specify the work they intend to delegate, recognise when its circumstances change and intervene with enough context to continue appropriately. Under substantial AI autonomy, the important skill is not writing an elaborate prompt. It is knowing what to hand over, what evidence to check and when to take the task back.
Heidi's September 2026 strategy statement, checked on 28 September 2026, separates removing clerical work from medical judgement, which remains with the practitioner. Its preview, checked on the same date, excludes the UK and Europe. This article considers the educational implications of substantial authorised delegation, not a Heidi-specific training standard or confirmed curriculum.
Specify the outcome, not just the document
Compare two fictional instructions. The first asks an agent to write a referral letter. The second defines the intended administrative outcome: prepare the agreed referral using the reviewed information, preserve uncertainty, present the proposed destination for confirmation and stop before transmission if the available records conflict.
The difference is not simply prompt length. The second instruction makes the clinician's intention and the agent's boundaries more inspectable. It distinguishes preparing content from acting externally and identifies a condition that requires human judgement.
A product should support that specification through sensible workflow design. Clinicians should not have to remember every possible control or repeatedly type information already available in the approved context. Nevertheless, they need to recognise when an instruction leaves an important decision unspecified.
"Arrange the appropriate follow-up" is a useful teaching example because it can conceal both clinical judgement and administration. Has the clinician already decided the purpose and timing, leaving only coordination? Or is the agent being asked to determine the plan? A learner should be able to explain that difference before discussing which button to press.
Separate repetitive handling from educational work
Not every manual task teaches medicine. Copying the same approved information between systems may add little educational value. But some activities contain reasoning even when they look administrative: selecting the relevant history, explaining why a referral is needed, reconciling conflicting information or recognising that an old plan no longer fits the current situation.
The educational question is where that reasoning occurs after automation. A trainee might formulate the clinical problem first and then delegate document handling. Alternatively, they might accept a polished summary without ever considering which details were important. Those are different learning arrangements even if the final document looks similar.
It would be a mistake to assume that automation necessarily causes deskilling. It would be equally weak to assume that time released automatically becomes better learning. Both claims require evidence about what trainees do, what feedback they receive and whether their performance transfers to situations without the tool.
The American Medical Association's March 2026 survey records interest in further AI training alongside concern about skill loss. Those are physicians' reported attitudes, not evidence that Heidi Work has improved or impaired competence. They support making the educational question explicit rather than treating onboarding as sufficient.
Test two competing learning hypotheses
Under the first hypothesis, delegated administration releases attention for clinical reasoning and patient communication. The learner spends less effort on handling and more on explaining the problem, comparing alternatives and discussing uncertainty with a supervisor.
Under the second, automation removes exposure to details through which the learner would otherwise understand the case. The trainee sees the completed output but misses the discrepancy that an experienced clinician would notice while constructing it.
A useful educational evaluation would assess both possibilities. Learners could explain a case before seeing the agent's proposed output, identify what additional information matters and justify when the workflow should stop. Later assessment could use a different case and an unavailable agent to examine whether the relevant skill has transferred.
Time on task alone would not distinguish the hypotheses. Faster completion might reflect improved organisation, superficial engagement or both. The outcome should include the quality of the learner's explanation and intervention decisions, not merely the elegance of the finished letter.
Rehearse the point where the workflow changes
Delegation scenarios can be taught without placing a real patient into an untested workflow. A proposed simulation might begin with an apparently straightforward administrative action and then introduce incomplete context, an ambiguous instruction or an unexpected response.
In one fictional exercise, a trainee authorises preparation of an agreed referral. A previous specialist letter then appears to describe a different plan. The assessment question is not whether the trainee can generate a more fluent referral. It is whether they recognise the conflict, recover the relevant context and pause the external action until the discrepancy has been considered.
In another exercise, the system reports that a follow-up request has been sent. The learner must distinguish that transmission from an accepted booking. In a third, the recipient's reply introduces a concern beyond the delegated administrative scope. The learner must recognise the change rather than continue to treat the interaction as routine coordination.
These are proposed teaching scenarios, not claims that Heidi Work or iatroX currently provides a dedicated library of Heidi delegation cases. Their purpose is to assess judgement around the workflow boundary, not a memorised script or prompt-writing contest.
Practise taking over, including when the agent is unavailable
A clinician taking over should be able to establish what was authorised, which actions have occurred and what remains unresolved. That reconstruction matters because repeating an external action may be less appropriate than completing the existing one.
A fallback exercise could remove the agent partway through a fictional workflow while leaving a record of its completed actions. The learner would then identify the current state, decide what information is missing and continue through the permitted manual process. The exercise should not reward confident improvisation when the available evidence is insufficient.
This is different from an onboarding tutorial. Knowing how to start a workflow does not demonstrate competence in recognising when it should stop. A meaningful assessment would include explanation, intervention and recovery, with feedback on the specific judgement that was sound or needed improvement.
The accompanying iatroX analysis of trust after an agent error explores the same issue from the adoption side: continuing to use a tool should depend on understanding its boundaries, not on maintaining confidence regardless of what happened.
Where clinical learning tools fit, and where they do not
This article is published by iatroX and includes its own learning tools in the discussion. The prospective task-execution role considered for Heidi Work and the educational role described here address different needs; neither category replaces the other merely because both use AI.
According to iatroX's September 2026 product information, Socratic Tutor opens on an attempted question, asks targeted follow-ups and works through the learner's misconception rather than simply supplying the answer. That design offers a way to practise explaining a distinction or revisiting a reasoning error. It is not Heidi-specific certification or evidence that a learner can safely operate any particular clinical agent.
Per iatroX launch information for September 2026, iatroX Simulations launched with 18 examination-specific tracks and 1,114 clinician-reviewed cases. Voice and text interaction, pause-and-coach practice, uninterrupted exam mode and transcript-linked feedback provide different ways to rehearse and review performance. Clinician review does not mean examining-body endorsement, and the platform is not presented here as predicting pass rates or replacing bedside and procedural practice.
For a learner who needs to check the basis of a general clinical proposition, Ask-iatroX is a different tool again. As described in September 2026 product information, it is a free reference grounded in NICE, CKS, SIGN and SmPC information from emc, with linked sources and a published retrieval and checking methodology. Source links help inspection; their presence is not proof that every output is correct or applicable to a particular patient.
Pay for a relevant learning pathway, not unrelated exam access
As published by iatroX for September 2026, Ask-iatroX and free question access have no trial expiry or verification gate. The paid subscription combines question banks, Socratic Tutor, the study planner, iatroX Simulations and CPD tools. Simulations and CPD are included, not separate add-ons.
The published price is £99 paid upfront for a year, equivalent to £8.25 per month billed annually, or £29 per month. Using those September 2026 prices, three monthly payments total £87 and four total £116, so the annual option is cheaper from the fourth month. The relevant value is several learning methods supporting one professional goal, not hypothetical access to examinations the reader will never sit.
For clinicians implementing a live agent, product-specific and locally approved workflow training remains the appropriate route to operating that system. For learners strengthening reasoning or communication, question-based tutoring and simulation can serve a complementary educational purpose. Professional learning records should remain distinct from claims of formally accredited CME, and neither a subscription nor a completion record establishes competence by itself.
Frequently asked questions
Will doctors mainly need better prompting skills to use Heidi Work?
Prompting may be part of the interface, but the more important proposed competencies are specifying outcomes, recognising exceptions and judging when to intervene. Product design should support those decisions rather than depend on elaborate free-text instructions.
Does greater automation necessarily cause clinical deskilling?
No, it could release attention for learning or reduce exposure to valuable reasoning work, depending on the task and educational arrangement. Those competing possibilities should be tested through performance and transfer, not inferred from time saved alone.
Does iatroX provide Heidi Work certification?
This article does not describe a Heidi-specific certification, integration or approved training programme. Its delegation exercises are educational proposals, while iatroX's described tutoring and simulation tools support broader clinical learning.
Practise explaining your clinical reasoning with iatroX Tutor →