When a patient does not resemble the population addressed by a recommendation, the clinician needs to assess applicability rather than simply retrieve more recommendations. Multimorbidity, frailty, treatment burden and personal priorities may change the balance. An AI-generated list of individually plausible actions is not necessarily a coherent plan for the person.
More guideline coverage does not automatically produce better care
A system may find relevant recommendations for every condition in a record and still fail to reconcile them. The resulting plan can be comprehensive in coverage but poorly prioritised, difficult to follow or insensitive to the patient's main concern.
NICE guideline NG56 on multimorbidity, published on 21 September 2016 and checked on 10 October 2026, addresses reducing treatment burden and considering how conditions and treatments interact. Its relevance here is the need for an individualised approach, not a special exemption from careful evidence use whenever a case becomes complex.
The question is not whether guidance should be followed or ignored. It is which recommendations apply, what they would mean together and how the patient understands the available choices.
A fictional patient with several reasonable plans
Consider a fictional older adult with heart disease, chronic kidney disease and painful arthritis. Different clinicians have suggested investigations, reviews and lifestyle changes. The patient is increasingly concerned about getting to appointments and wants enough comfort and energy to continue an important family activity.
An AI summary could organise each condition accurately and produce a separate management section for each. Yet the combined document might leave the central questions unanswered: which action matters most now, which proposals conflict, what can realistically be undertaken and who will coordinate the plan?
The educational exercise is not to invent a single correct treatment from this incomplete description. It is to identify what further information is needed. The patient's current symptoms, previous responses, medicines, functional priorities and existing specialist plans all affect the discussion.
A useful output would make those gaps visible. A less useful output would convert them into assumptions and present a longer plan as though length demonstrated individualisation.
Treatment burden belongs in the clinical discussion
An appointment is not only a line in a plan. It may involve travel, time away from responsibilities, preparation, assistance and another set of instructions. These practical demands can change what a patient is able or willing to undertake.
A UK primary-care cross-sectional study published in 2021 examined treatment burden among people with multimorbidity and associations with factors including health literacy and financial difficulty. Those associations do not prove a particular AI plan causes poor outcomes, but they support asking about capacity rather than treating a plan's practical demands as irrelevant.
For the fictional patient, the clinician might explore which activities are most difficult, which appointments could be coordinated and what the patient believes each intervention is intended to achieve. The goal is not to reduce care indiscriminately. It is to make the agreed care purposeful and workable.
An AI assistant can help organise those questions. It cannot know the answers merely because the electronic record contains several diagnoses.
Frailty is not permission to assume preferences
Frailty, age and a long problem list should not become shortcuts for inferring what someone values or what treatment they would decline. Two people with similar diagnoses can have different goals, support and tolerance for the burdens of treatment.
The relevant conversation is about the individual, not a demographic stereotype. The clinician should distinguish a clinical judgement about likely benefit from an assumption about what the patient wants.
For an AI-generated plan, inspect statements such as "the patient would prefer conservative care" or "intensive follow-up is unsuitable". Are they based on an actual discussion, a justified clinical assessment or an inference made from a label? If the source does not establish the preference, the plan should not present it as established.
This is especially important in summaries, where a tentative consideration can become a concise but unjustified instruction for the next professional.
Sparse evidence calls for bounded conclusions
Some patients differ from the studied population in ways that make extrapolation uncertain. Adding more citations does not necessarily resolve that uncertainty if the references concern the same narrow population.
A useful answer should explain the limit precisely. The uncertainty might concern whether an effect applies, whether harms differ, or whether the proposed pathway can be delivered. Those are more informative statements than a generic warning that "clinical judgement is required".
For teaching, ask the learner to complete the sentence: "This recommendation may not transfer directly because..." Then require them to name the missing evidence or patient information that would make the decision clearer. The exercise encourages explicit reasoning rather than ritual caution.
The appropriate next step may be specialist discussion, review of original evidence, further assessment or a shared decision under acknowledged uncertainty. The answer should depend on the problem, not a universal rule that uncommon cases always need more testing.
Reconcile the plan before adding another action
A proposed review can begin with what is already happening. Which treatments and investigations are active? Which were only suggested? Which clinician owns the follow-up? Does the patient understand the reason for the current plan?
Only then consider additions. An apparently beneficial new action may duplicate an existing one or depend on information another team is already gathering. Conversely, a decision to simplify needs to consider what benefit might be lost and how the outcome will be reviewed.
The resulting plan should contain priorities, responsibilities and review points, not merely a catalogue of recommendations. Where professionals disagree, the disagreement should be made explicit and resolved through an appropriate clinical process rather than hidden by an AI-generated compromise.
Use structured questioning without mistaking it for validation
In iatroX's October 2026 product positioning, Brainstorm provides a setting for structuring clinical questions and alternatives. It can be used as an educational format for asking what has been assumed, what is missing and what might change the interpretation. It is not established protection against diagnostic error or a replacement for supervised decision-making.
A fictional learning exercise could ask the learner to form an initial problem representation, identify the competing priorities and then compare their questions with those suggested by the tool. The valuable output is a clearer account of uncertainty, not automatic agreement with the generated plan.
For the patient with several conditions, better reasoning may lead to a shorter plan, a different sequence or a more explicit conversation. It may also confirm that substantial treatment remains appropriate. The measure of quality is whether the plan fits the evidence and the person, not whether the AI produced the greatest number of actions.
Frequently asked questions
Does multimorbidity mean disease-specific guidelines should be ignored?
No. Their relevance, interactions and combined burden should be assessed within an individualised plan rather than applied as disconnected instructions.
Can AI infer a patient's treatment preferences from age or frailty?
It should not present such an inference as an established preference. The patient's own account and an appropriate clinical discussion remain necessary.
Can structured AI questioning replace specialist advice in a complex case?
No. It may help identify questions for review, but uncertainty requiring specialist assessment is not resolved simply by generating a structured answer.
Practise identifying the assumptions in a complex clinical case →