No single participant should silently decide what "local" means for every medical AI answer. Clinical evidence, national guidance, service arrangements and a patient's circumstances address different questions. A useful system should make their roles visible, explain unresolved disagreement and identify who has authority to approve a local adaptation.
The problem is practical. An answer can cite a national recommendation, mention a local service and still conceal the judgement that connected them. Did the system find an authorised pathway, infer what probably happens in that country or treat the nearest available document as the correct one?
This article proposes a governance approach, not a description of a completed OpenEvidence, Anthropic or iatroX implementation. The underlying principle has an established foundation: WHO's guidance announced on 18 January 2024 calls for early, structured involvement from healthcare professionals, patients and other stakeholders in the design of these systems. The decision rights and interface examples below are iatroX's proposals for putting that involvement into practice.
Start by asking which question a source can answer
A research paper may help establish the benefits and limitations of an intervention in a studied population. A national guideline may interpret a body of evidence for a defined jurisdiction. A local pathway may explain how a service accepts referrals. A specialist may identify why an unusual case falls outside a routine pathway.
These are not interchangeable forms of authority. The local referral document can be authoritative about the route into that service without establishing that every person outside its acceptance criteria has no clinical need. A recent paper can identify a relevant finding without automatically replacing an applicable guideline or resolving implementation.
A proposed clinical AI interface should therefore distinguish evidence about what may be beneficial, recommendations about what should ordinarily be considered, and operational information about how care is organised. Combining them into a single answer without preserving those distinctions makes disagreement harder to inspect.
The question is not simply, "Which source ranks highest?" It is, "Which source is relevant to this part of the decision, and what does it leave unanswered?"
A country label cannot establish local authority
A national setting may be a useful starting point, but it cannot identify every organisation, service or professional context within that setting. Nor does selecting a country establish that the patient fits the population covered by a retrieved document.
For a fictional regional clinic, the relevant context might include its referral network, the type of appointment, which results the clinician can access and the service responsible for follow-up. Those details should come from identifiable information, not assumptions made from an address or language choice.
Even a named organisation is insufficient when a document applies only to a particular service. A pathway for one specialty should not become a default rule for the whole organisation merely because both share a logo.
The useful design challenge is to record the scope of authority alongside the source. Who issued the document? Which service and population does it cover? Is it operational guidance, a clinical recommendation or an educational explanation? When was that scope last confirmed?
Distinguish three kinds of apparent conflict
Some disagreements disappear when the scope is examined. Two documents may address different populations, settings or stages of care. A system should not announce a controversy when it has actually retrieved answers to different questions.
Other disagreements arise from timing. An older local document may not reflect a newer national recommendation. The answer should expose that version difference without assuming that the newest document automatically settles every operational detail.
A third category is genuine disagreement or an unresolved implementation gap. Relevant sources may make different recommendations, or a service may be unable to deliver the pathway described in the evidence. That conflict should remain visible until an appropriately responsible person resolves it.
This classification is a proposed review aid. It does not authorise software to adjudicate professional disagreements independently. Its purpose is to make the next human judgement more informed: clarify scope, verify currency or seek a clinical or operational decision.
A fictional conflicting-recommendations panel
Imagine an educational case involving a non-urgent specialist assessment. The national document in the exercise describes a broad group for whom assessment should be considered. A local service document uses a narrower acceptance route. A specialist commentary discusses exceptions, while the fictional patient has concerns about travel and the proposed appointment format.
These documents and circumstances are invented solely to illustrate an interface. They are not a clinical pathway or a recommendation about any actual condition.
| What the proposed panel shows | Why it matters |
|---|---|
| National recommendation: the educational scenario falls within the document's stated population | Establishes the relevant recommendation without claiming that local access has been arranged |
| Local pathway: its acceptance route does not clearly cover this scenario | Identifies a service-routing gap rather than concluding that assessment is unnecessary |
| Specialist commentary: discusses exceptions but is not the local service's acceptance policy | Prevents an opinion from being presented as an authorised booking rule |
| Patient context: travel and appointment-format concerns were supplied in the case | Makes the person's circumstances visible without treating them as a clinical contraindication |
| Unresolved question: who can clarify the appropriate route for this scenario? | Directs attention to the missing decision rather than generating an artificial consensus |
A weak response would flatten these inputs into a confident instruction. A better response would explain which parts are established, where the documents differ and what remains to be clarified through the relevant professional or service route.
The panel should also make disagreement inspectable at the claim level. A citation at the end of a paragraph is insufficient when one source supports clinical consideration and another only describes administrative acceptance.
Participation needs decision rights, not just invitations
A workshop can reveal useful problems, but participation is limited if the people attending cannot influence what happens next. Meaningful local involvement requires a defined route from a concern to a decision.
For an implementation team, that could mean specifying who can propose a source, who reviews its scope, who approves an adaptation and who can suspend an unsafe local configuration. It should also be clear who responds when clinicians report a recurring mismatch between an answer and the service they actually use.
The following allocation is an example for discussion, not a universal governance structure or a description of either company's arrangements.
| Participant | Possible responsibility in a locally agreed process |
|---|---|
| Clinical content lead | Reviews clinical interpretation, scope and material disagreements |
| Local source owner | Confirms the document's provenance, currency and authorised use |
| Service operations lead | Confirms routing information, ownership and practical implementation |
| Product supplier | Implements approved changes, records versions and investigates system behaviour |
| Front-line clinicians | Identify mismatches, test realistic cases and challenge unclear outputs |
| Patient or public contributors | Examine whether the proposed information and workflow address relevant needs and burdens |
Responsibility should not be confused with unilateral control. A supplier may be responsible for implementing a configuration while a clinical team remains responsible for approving its interpretation. A service manager can clarify availability without independently resolving a clinical evidence dispute.
Build a correction route that closes the loop
A feedback button is only the beginning. A useful correction process needs to preserve what was reported, the relevant answer and source versions, the nature of the concern and the decision taken.
For a fictional report, a clinician might say that the system repeatedly selects a pathway belonging to the wrong service. The investigation should distinguish a retrieval error from inaccurate document metadata, ambiguous user context or a source that was never appropriate for the intended audience.
The resulting action could differ substantially. Correcting the document's scope is not the same as changing a prompt. Removing obsolete content is not the same as rewriting a valid recommendation. Each action should be tested against related cases so that fixing one mismatch does not introduce another.
Closing the loop also means telling the reporting user what happened where appropriate. "Feedback received" is not equivalent to "the source was reviewed and the affected configuration corrected". The distinction matters for trust and for whether users keep reporting problems.
Patient involvement is not a substitute for clinical evidence
A patient contributor should not be expected to adjudicate the statistical validity of a study. Their contribution can be different and equally relevant to the usefulness of the product: identifying unclear explanations, impractical assumptions, access burdens or information that people need before discussing an option.
Similarly, a person's preferences should not disappear because a tool can generate a technically coherent pathway. In the fictional panel, concern about travel does not prove that an assessment is unnecessary. It identifies something that needs discussion when deciding how a suitable plan can be implemented.
For product design, this suggests separating the evidence statement from the contextual discussion. The system can show what a source supports while making clear that a patient's circumstances were supplied, not inferred, and may change which practical questions need resolving.
That is more transparent than presenting a personalised-sounding answer whose underlying assumptions the person cannot see or correct.
Keep an adaptation history, not just a current answer
Local ownership becomes difficult to inspect if the only visible object is the latest generated response. A proposed adaptation record should preserve the source, the reason for the change, the people or roles approving it, its intended scope and the conditions that should trigger review.
The record need not expose confidential internal discussions. It should make material changes understandable. A reader should be able to distinguish a national recommendation reproduced without modification from a local interpretation or an operational instruction added to support a particular service.
An expiry or review date alone is insufficient. A document can become unsuitable before its planned review if the service changes. Conversely, a scheduled review date does not establish that a recommendation has become clinically wrong. The relevant question is what has changed and who has assessed the consequences.
These proposals complement iatroX's separate articles on WHO SMART Guidelines and the proposed context receipt. Source structure, visible assumptions and governance answer connected but different questions.
The same standard should apply to iatroX
Per iatroX product information, September 2026, Ask-iatroX is grounded in NICE, CKS, SIGN and SmPC information from emc. Its published methodology describes source retrieval, ranking, citation grounding, output checking and uncertainty handling.
Those are design features, not proof that every answer correctly resolves local disagreement. A UK-focused source collection does not establish that the system knows a particular service's current arrangements or the circumstances of an individual patient.
The standard proposed here should therefore challenge iatroX as well as larger platforms: make the relevant jurisdiction inspectable, avoid inventing local detail and preserve uncertainty where a source does not settle the question. The decision-rights process and conflict panel in this article are proposals, not claims that these exact interfaces are already implemented.
Local medical AI becomes more useful when its authority can be examined. The goal is not to eliminate professional disagreement through fluent wording, but to show clinicians what is known, what is assumed and who needs to make the next judgement.
Frequently asked questions
Should a medical AI tool always prioritise local guidance over international evidence?
No: the appropriate source depends on the question, its scope and its currency, and an operational pathway does not automatically resolve a clinical evidence issue. A useful answer should explain the relevant difference rather than apply an invisible universal hierarchy.
What would count as meaningful local clinician involvement?
Clinicians should have a defined route to influence source selection, challenge adaptations, test realistic cases and obtain a response to reported problems. Attendance at a consultation exercise alone does not establish those decision rights.
Does the Anthropic and OpenEvidence announcement establish this governance model?
No: the decision-rights framework and fictional conflict panel are proposals developed in this article. They are questions to apply to locally adapted tools, including iatroX, rather than claims about an undisclosed implementation.