A medical AI answer can accurately describe an evidence-based pathway and still fail to help the clinician who asked the question. The missing element may be access to an investigation, the timing of specialist advice, the patient's ability to attend or an uncertainty the system has quietly treated as resolved. Correctness and usability overlap, but they are not identical.
The current interest in locally adapted clinical AI makes this distinction important. iatroX's main Anthropic and OpenEvidence partnership explainer covers the September 2026 announcement. This article develops a clinical design argument: useful adaptation should expose constraints without silently lowering the standard of care.
All scenarios below are fictional educational examples. They are not treatment protocols, descriptions of particular health systems or observations from a comparative product test.
An answer has more than one kind of correctness
There is factual correctness: does the answer represent the evidence accurately? There is applicability: does the cited material address the relevant patient group and clinical question? There is operational usefulness: can the clinician understand what the answer requires in this setting?
A failure in any of these areas can matter. An accessible intervention is not useful if its rationale is wrong. An accurate recommendation is incomplete if its conditions are omitted. A well-supported pathway may still leave the user uncertain when a necessary resource is unavailable.
The proposed response is not to replace evidence with whatever happens to be convenient. It is to make the relationship between evidence and circumstances explicit. The output should identify what the evidence supports, what the system actually knows about the setting and what remains unresolved.
WHO's January 2024 guidance on large multi-modal models warns that plausible outputs can be inaccurate or incomplete and that users may over-rely on them. The practical concern here is a particular form of incompleteness: an answer that hides the conditions required to use it.
Scenario one: the investigation exists, but not in this consultation
A fictional community clinician asks an educational question about a pathway that includes an investigation performed elsewhere. The AI describes the investigation accurately and finishes with "arrange the test and review the result".
That may identify part of the pathway without answering the user's practical question. The clinician might need to establish the referral route, whether the investigation's timing affects the decision, who will receive the result and what information is missing before urgency can be assessed.
The response should not invent those details. Nor should it assume that absence from the current premises means the investigation is unavailable throughout the area. "Not available here", "not available today", "not accessible to this patient" and "availability unknown" are different statements.
A resource-aware format would preserve the investigation's role while exposing the unresolved dependency. It would explain that the supplied information does not establish the access route or acceptable timing, rather than replacing a necessary step with a convenient but unsupported alternative.
Scenario two: a specialist is not immediately reachable
In a fictional case-discussion exercise, a clinician has a question whose interpretation may require specialist input. The specialist service exists, but the user has not established when advice can be obtained.
An unhelpful answer can fail in either direction. It may write "consult the specialist" without recognising the gap between identifying a need and obtaining advice. Alternatively, it may generate a confident substitute plan, treating the lack of immediate access as permission to resolve uncertainty on its own.
A better structure distinguishes the reason specialist input matters from the practical uncertainty about obtaining it. It identifies which facts would change the assessment and makes clear that a response cannot determine a safe delay from a description that does not contain the necessary information.
This does not require the AI to invent a local telephone number or an escalation pathway. Where an approved local route is available, it can be cited. Where it is not known, the uncertainty should be visible rather than hidden behind generic reassurance.
Scenario three: an option cannot be obtained by the patient
A fictional learner considers a management discussion in which a recommended option is not accessible to the patient. The obstacle could be local availability, travel, affordability or a preference that has not yet been explored. These constraints are not interchangeable.
A response should not translate "cannot currently obtain" into "does not need". It should also avoid assuming that a different option is equivalent because it is easier to access. Equivalence is an evidence question, not a convenience label.
The educational task is to identify what requires checking: the evidence for alternatives, the patient's priorities, the consequences of delay and the appropriate clinical discussion. No alternative treatment is recommended in this scenario. The point is that the missing information must remain part of the problem rather than disappear when the answer is generated.
A before-and-after example of response design
The following comparison is an original interface-writing example. It is not an output captured from OpenEvidence, iatroX or another product.
| Generic response pattern | Context-aware response pattern |
|---|---|
| "The guideline recommends an investigation, followed by specialist review." | "The cited pathway includes an investigation and specialist input. Whether those steps apply here depends on the patient information and pathway conditions." |
| "Arrange the investigation locally." | "The user has said the investigation is not available at the current site; access elsewhere has not been established." |
| "Use an alternative if necessary." | "An alternative cannot be assumed equivalent. Its evidence, suitability and access need checking before it can be compared." |
| "Follow up when results are available." | "Ownership of follow-up and the implications of waiting have not been supplied. This answer does not establish a safe interval or a completed plan." |
The improved format is not better because it is longer. It is better specified: it separates a source-supported statement from a setting-dependent assumption. In a real interface, the aim should be a compact answer that makes the important distinction visible without overwhelming the user.
The three-part answer a clinician can inspect
The first part is what the evidence supports. This should identify the relevant source, population and recommendation, including important conditions or exclusions. It should not claim to describe a local service merely because it cites a national guideline.
The second part is what has been established about the setting. User-supplied information belongs here, alongside verified institutional information where the system is authorised to use it. The origin and date of a setting fact matter. A pathway document may describe a service in principle without proving that it is operating normally at this moment.
The third part is what remains unresolved. This includes missing clinical information, uncertain resource access, conflicting guidance and circumstances requiring an appropriate local or specialist decision. The answer should explain which uncertainty is material rather than list every imaginable caveat.
iatroX's proposed context-receipt article develops a compact way to display these distinctions. The receipt is a design proposal, not an existing certification or a claim that every current platform implements it.
Adapting is not the same as compromising
There is a risk that resource awareness becomes a euphemism for accepting a lower standard without saying so. A system might learn to give different answers to different country labels even when the actual clinical and resource information is identical.
That is not the intended goal. The response should change because a relevant, supported fact changes, not because the interface associates a location with a vague assumption about what care is possible.
An unavailable resource may create an unresolved risk or a need for an alternative pathway. It does not make the original clinical need disappear. A transparent answer should show that distinction and avoid implying that the remaining option is equivalent unless the evidence supports that conclusion.
The reverse problem also matters. A system can be unusably rigid if it repeats a pathway without acknowledging a verified constraint. Good adaptation therefore requires both fidelity to evidence and honest treatment of the setting, rather than choosing one and ignoring the other.
Test the change in context, not just the initial answer
A practical evaluation could present the same fictional clinical information with different resource conditions. In one version, the relevant investigation is available through a documented pathway. In another, availability is unknown. In a third, a verified constraint prevents access through that route.
The test is not whether the wording changes. It is whether the change is appropriate. Does the system preserve the clinical rationale? Does it invent an alternative? Does it identify the missing information that now matters? Does it become more certain merely because the user asks for a simple answer?
Locally relevant clinicians should help define acceptable responses before seeing the outputs. An assessment should record both unsafe overconfidence and unhelpful refusal, rather than reward a system for declining every difficult case. The companion article proposing a context-sensitive benchmark develops this method; no benchmark results are claimed here.
What this means for clinical learning
Resource constraints make useful educational variables because they force the learner to distinguish evidence from assumptions. A case can ask why a recommendation applies, what information would change the decision and what cannot safely be inferred from the available setting description.
Per iatroX product information, September 2026, its simulations offer voice and text practice, a pause-and-coach practice mode and an uninterrupted examination mode, with transcript-linked feedback and Tutor-led remediation. These are learning features, not a claim that simulated performance establishes competence to manage every resource-constrained clinical situation.
Clinician review of a case also has a defined scope. It is not the same as an examining body's endorsement or independent validation of every generated interaction. Simulation should sit alongside supervised practice, source checking and appropriate local training.
A productive learning goal is not "the chatbot found a way around the shortage". It is "the learner recognised the constraint, preserved the clinical concern and explained what still required resolution". That is a more credible bridge from AI-assisted answers to professional judgement.
Frequently asked questions
Can an evidence-based AI answer still be inappropriate for a local setting?
Yes: a recommendation may depend on resources, pathways or patient circumstances that have not been established. The answer should distinguish those dependencies from the evidence itself.
Should medical AI recommend a different standard when resources are limited?
It should not silently redefine necessary care as unnecessary or assume that a convenient alternative is equivalent. Resource constraints should be made explicit, together with the uncertainty and decisions they create.
Can simulations teach reasoning about unavailable resources?
They can provide structured practice in recognising constraints and explaining what remains unresolved, but that is an educational design opportunity rather than proof of clinical competence. No simulation removes the need for supervised practice or appropriate local guidance.
Structure a clinical learning question with iatroX Brainstorm →