An agent does not need to make the final diagnosis to influence clinical care. Selecting what belongs in a referral, deciding which history deserves prominence or categorising an unresolved issue can shape the information on which a clinician acts. Calling the final operation administrative does not make every preceding interpretation clinically neutral.
Heidi's 28 September 2026 announcement places visit preparation, follow-up work and actions across practice systems within Heidi II's proposed scope. This article examines where interpretation could enter such workflows. It is not a claim that Heidi II is an autonomous diagnostic service or a legal conclusion about its regulatory classification.
Copying information is different from deciding what matters
Consider two fictional tasks. In the first, a clinician has approved a document and asks for it to be transferred to an identified destination without changing its content. In the second, a system selects relevant information from the record, decides what to include and prepares the document for approval.
The final visible step might be identical: sending a letter. The intellectual work is not. The second workflow has made choices about relevance, chronology and emphasis before the clinician reaches the approval stage.
Those choices may be helpful. An organised draft could reveal a missing detail or make a complicated history easier to understand. The point is not that selection is inherently unsafe. It is that the scope of evaluation should include selection, not stop at whether the letter arrived.
The referral context can change the receiving team's understanding
Imagine a fictional referral whose source material contains both an established problem and a recent change in the patient's circumstances. A concise draft accurately describes the established problem but gives little prominence to the change. Another draft uses the same facts but makes the new development central.
Neither document needs to contain a fabricated statement for their implications to differ. The receiving team may understand the reason for referral differently because the material has been organised differently. The clinician reviewing the draft therefore needs to assess the clinical account, not only the spelling and recipient.
This is a scenario to test, not an allegation about a Heidi-generated referral. It illustrates why a workflow that appears administrative at its endpoint may contain an earlier judgement about what the patient needs another clinician to know.
Recall eligibility contains assumptions too
A proposed recall process may begin by looking for missing records of a completed action. That can be useful, but missing documentation is not necessarily proof that the action is outstanding. The patient may have completed it elsewhere, a record may not yet be reconciled or an individual decision may change what should happen next.
The administrative step is creating or sending the recall. Before that comes an interpretation of eligibility. A different interpretation of the same information could change who is contacted or excluded.
We would therefore expect an evaluation to inspect both the selected group and the patients left out. A convincing-looking list does not establish that its rule was appropriate. Equally, an existing manual process should be assessed against the same standard rather than assumed correct because a person produced it.
Categorising unresolved work is a clinical influence point
Another fictional workflow sorts outstanding issues into categories for later review. A note describing an uncertain plan is placed under routine administration because it contains the word "follow-up". The task itself does not diagnose anything, but the categorisation could influence when and by whom the underlying issue is considered.
The relevant assessment is whether the category preserves the uncertainty and context that should determine the next human decision. Where those details are absent, the system should not silently turn an ambiguous issue into a settled administrative instruction.
This suggests a broader design question: can the clinician see why the item was categorised that way? A label without its basis may be efficient to scan but difficult to challenge. A useful interface would connect the classification to the material facts and unresolved assumptions.
Use a boundary test based on consequences
A practical test is to ask: could a different interpretation of the same information change this patient's care? When the answer is yes, the workflow contains a clinically consequential interpretation even if the final action is administrative.
That does not determine legal status. It helps identify what must remain inspectable and what should be evaluated. It also avoids an unhelpful division in which every task is either entirely clerical or equivalent to an autonomous doctor.
The next questions are specific. Which facts drove the interpretation? Were competing explanations preserved? Did the clinician explicitly supply the intended action, or did the system infer it? Can the proposed work be corrected without leaving dependent tasks attached to the old assumption?
Make invisible decisions visible before approval
A proposed review process would distinguish source information, interpretation and authorised instruction. It would show which parts came directly from the record, which were supplied by the clinician and which were generated by the system. The goal is not to display every technical operation, but to expose the choices that could change the result.
In the referral scenario, that might mean making the reason for prioritising a particular detail inspectable. In the recall scenario, it could mean showing why the patient qualifies and which uncertainties remain. In the unresolved-work scenario, it would mean explaining the chosen category without presenting it as an established clinical fact.
These are proposed transparency requirements, not verified Heidi II interface features. Heidi's 29 September 2026 trust discussion describes clinician control and escalation when tasks are unclear. A buyer should examine how those principles apply to the particular interpretation points in its own workflow.
Evaluate influence without assuming harm
An approved simulation could compare source records with the summaries, categories and tasks created from them. Independent reviewers would identify whether any change in emphasis altered the apparent clinical question. The assessment should include examples where automated organisation improves the account as well as examples where it narrows it inappropriately.
Useful outcomes would include recognition of relevant changes, preservation of uncertainty and justified correction by the reviewer. A high rate of document delivery would not answer those questions, although delivery remains a separate operational requirement.
No such evaluation has been run for this article. The purpose is to define what needs testing. As of the September 2026 launch, Heidi II's new capabilities were not available in the UK or EU, so international capability claims should not be treated as evidence of local deployment.
The same boundary applies to iatroX: selecting sources and presenting an answer can influence interpretation, even where the product does not execute a workflow. Clear sources and explicit limits help scrutiny; they do not remove the need for it.
Frequently asked questions
Can administrative AI influence a clinical decision?
Yes, it can do so by selecting, emphasising or categorising information before a clinician makes the final decision. Whether that influence is beneficial depends on the task and implementation.
Does this mean Heidi II is an autonomous diagnostic system?
No, clinical influence and autonomous diagnosis are different claims. This article analyses workflow interpretation rather than asserting an undocumented diagnostic function or legal classification.
What is a useful boundary test for a proposed workflow?
Ask whether a different interpretation of the same information could change the patient's care. Where it could, make the consequential assumptions visible and evaluate how they are reviewed.