Doctors should consider offloading repetitive handling while retaining the interpretation needed to recognise contradictions, judge applicability and respond when the task changes. Information retrieval sits between those categories: assistance can be useful, but the retrieved material still needs appraisal. Heidi II makes this selective division more important than a blanket choice between using AI and doing everything manually.
The product context is Heidi's September 2026 announcement, which describes Agents, Memory and Routines around clinical work. The new capabilities are not available in the UK or EU. The delegation framework below is an analytical proposal, not a measured account of how Heidi II changes doctors' thinking.
Separate handling, retrieval and interpretation
Mechanical handling includes moving an approved document, locating a known attachment or placing agreed information in the appropriate field. These tasks require accuracy and permission, but they need not involve deciding what the patient's problem means.
Retrieval involves identifying relevant information. It can save effort, yet relevance depends on the question. A useful source for a general explanation may not settle an individual case, and a recent document may not supersede an older one on every point.
Patient-specific interpretation involves deciding how information fits together, what remains uncertain and whether an intended action is still appropriate. These categories can occur within the same workflow. The clinician should not assume that delegating a mechanical endpoint means no interpretation happened beforehand.
The positive scenario: less handling, more attention
Imagine a fictional clinician seeing a patient with several ongoing conditions and a complicated sequence of appointments. Before the visit, assistance organises the correspondence and identifies where expected information is missing. The clinician uses the released attention to hear the patient's priorities and understand what has changed.
In this scenario, offloading has not removed judgement. It has reduced the work needed to reach the information on which judgement depends. The clinician may become better informed because relevant material is easier to inspect.
That is a plausible benefit worth measuring. An evaluation should look beyond whether the preparation was faster and ask whether the consultation became more attentive, whether unresolved questions were recognised and whether subsequent work matched the agreed plan. It should not assume those benefits merely because less typing occurred.
The concerning scenario: an interpretation arrives ready to approve
Now change the workflow. The clinician receives a coherent account of the same patient, accepts its priorities and authorises the proposed follow-up without constructing a separate understanding of the unresolved issues.
The clinician may still agree correctly. The concern is that agreement has become harder to distinguish from recognition of a plausible story. If the patient introduces a contradiction, the doctor may need to reconstruct context that they never independently assembled.
This is not evidence that Heidi II causes passive approval. It identifies the effort whose disappearance would matter: not remembering every detail, but understanding which details could change the task. A product evaluation should distinguish that loss from the removal of genuinely repetitive work.
What the knowledge-worker research says
A CHI 2025 study surveyed 319 knowledge workers about generative AI and critical thinking. Greater confidence in AI was associated with less reported critical thinking, while respondents also described a shift towards verification and integration. These were self-reports from non-clinical work, not measurements of doctors' performance, neurological harm or inevitable skill decline. The source is the original research publication.
Its relevance is the possibility that effort moves rather than simply disappears. A doctor may spend less time producing an account and more time verifying it. Whether that is beneficial depends on the quality of verification and whether the doctor can still identify what deserves scrutiny.
The study does not establish that lower reported effort is intrinsically bad. Removing unnecessary effort can be the purpose of a useful tool. The question is whether the remaining effort supports the responsibility the user still holds.
Delegate the operation, retain the uncertainty
A selective approach would begin by identifying the uncertainty attached to each task. If the clinician has already approved the content and destination, delegation may concern execution. If the system must infer whether the action is needed, the review should include that inference.
For the fictional patient, an agent might organise appointment correspondence while the clinician retains responsibility for reconciling a changed preference with an earlier plan. The clinician need not remember every administrative detail, but should know which issue remains unresolved and what information would settle it.
We would also distinguish a stable rule from a patient-specific exception. A routine can follow an agreed process while still needing a clear stopping point when a case falls outside its assumptions. The quality of that stopping behaviour belongs in the evaluation, not only the speed of the routine when everything fits.
Make selective delegation workable under pressure
A framework that requires lengthy deliberation before every action will not help a busy service. The practical aim is to identify recurring boundaries in advance. A practice could define which tasks are mechanical, which require review of selected assumptions and which remain dependent on an explicit clinical decision.
The user should be able to recognise when a task crosses a boundary. A missing attachment is an operational problem; a new patient concern may change the task itself. Those situations need different responses even when both appear as incomplete work in the interface.
Training should therefore focus on recognising the transition, not on memorising an exhaustive list of clicks. A colleague taking over should understand the original objective, the unresolved issue and the permitted next step without needing the initiating clinician's private account of the case.
Evaluate what clinicians keep doing well
A proposed evaluation could observe how users allocate attention before and after introducing assistance. It would compare their ability to identify consequential contradictions, explain the plan and recover an interrupted workflow, alongside remaining administrative effort.
Think-aloud exercises may help reveal the reasoning in simulations, while routine operational data can describe completed work. Neither should be treated as a complete account on its own. Self-reported ease, task quality and retained competence are distinct measures.
No cognitive-offloading study of Heidi II has been run for this article. The purpose is to define a useful assessment rather than predict an outcome. A constructive result would show that the product removes low-value handling while preserving or improving the understanding required for care.
The same principle applies to reference and educational tools, including iatroX. A convenient answer can reduce retrieval effort, but the clinician still needs to understand the question, inspect the relevant evidence and recognise where the answer does not resolve the case.
Frequently asked questions
Is cognitive offloading inherently harmful for doctors?
No, reducing repetitive work can be useful and may release attention for patient care. The important distinction is which effort is removed and which judgement must remain dependable.
Did the CHI study show that AI damages clinicians' thinking?
No, the 2025 study concerned knowledge workers' self-reports and associations between confidence and reported critical thinking. It did not study Heidi II, clinical performance or neurological harm.
What should remain with the clinician when a task is delegated?
The clinician should retain enough understanding to assess consequential assumptions, recognise changed circumstances and intervene appropriately. That does not require manually reproducing every technical step.