Clinicians' management plans were rated more favourably than AMIE's for practicality and cost-effectiveness in the exploratory analysis of its primary-care feasibility study. These were assessments of plans, not measurements of actual healthcare spending or proof that AMIE routinely overinvestigated. The finding appears in the March 2026 preprint and the researchers' account updated on 8 October 2026.
It is a useful counterweight to coverage centred on diagnostic lists. A plan can contain recognisable medical knowledge yet leave important questions unanswered: what should happen first, what the patient can realistically do and who will review the outcome.
What the comparison actually assessed
The preprint describes blinded clinical evaluation of differential diagnoses and management plans. Some dimensions did not show a statistically significant difference, while practicality and cost-effectiveness favoured the clinical plans. The March 2026 methods and results provide the detail.
The plans being rated were not all delivered as competing care pathways. AMIE's research management plans were stored separately and were not shown to patients or their clinicians. This makes the result a judgement about the assessed outputs, rather than a trial demonstrating lower expenditure or better outcomes after implementation.
The distinction protects the favourable and unfavourable findings equally. A high rating is not proof that a plan worked for a patient. A lower practicality rating does not establish that the system caused unnecessary investigations. Both would require evidence about what actually happened.
Two fictional plans can contain similar knowledge
Consider a constructed case involving knee pain that is affecting a person's ability to work. The scenario deliberately omits the information needed for an actual clinical recommendation; it is an exercise in organising a plan.
The first fictional plan lists possible investigations, exercise options, several treatment categories, referral possibilities and a suggestion to review. Much of the content may be relevant in some circumstances, but the plan does not say which question must be settled first, what the patient prefers or who will arrange follow-up.
The second fictional plan begins by clarifying the symptom pattern, functional impact and findings that would change urgency. It then separates what can be agreed after assessment from what depends on further information. It identifies the patient's main practical obstacle and makes the review responsibility explicit.
The exercise does not declare the second plan medically correct from the sparse facts supplied. It demonstrates a structural difference: one presents a menu, while the other makes the decision sequence and its dependencies visible. Neither is an observed AMIE or clinician output from the study.
What makes a plan executable?
A practical plan connects a clinical objective to an action that someone can carry out. It distinguishes an immediate decision from a possibility to reconsider and makes uncertainty visible rather than hiding it behind a long list of options.
| Element | Question a reviewer should be able to answer |
|---|---|
| Priority | What problem is being addressed first, and why? |
| Dependency | Which decision cannot be made until another fact is established? |
| Patient fit | What preference or practical constraint changes the approach? |
| Local fit | Can the proposed service or investigation be accessed through the relevant pathway? |
| Ownership | Who is responsible for the next action and for checking its result? |
| Reassessment | What would cause the plan to change? |
This is an editorial review framework, not the study's original scoring rubric. It translates the broad idea of practicality into questions that can be used in teaching or a future evaluation.
For a UK clinician, the patient-fit component also connects to the GMC's decision-making and consent guidance, checked on 10 October 2026. A plan should support a dialogue about relevant options and what matters to the person, rather than assume that technical plausibility settles the decision.
Why the clinical plans may have differed
The treating clinician had access to information and context that were not identical to AMIE's pre-visit conversation. Examination, records and the ordinary appointment could affect the plan. The March 2026 report discusses this asymmetry.
It is plausible that familiarity with local services, continuity or knowledge of the patient's preferences could contribute to practical planning. Those are interpretations to investigate, not explanations proven by the rating difference. The study does not allow each favourable rating to be assigned to one particular human capability.
It is equally possible that a different AI configuration, better contextual input or a more constrained planning task could change the comparison. That possibility does not negate the result. It identifies the next design and evaluation question.
Cost-effectiveness ratings are not a health-economic analysis
A clinician evaluating a plan can judge whether its proposed activity appears proportionate. That judgement is valuable, but it is different from observing costs, outcomes and trade-offs across a service.
A formal evaluation of economic value would need to establish the relevant costs, the outcomes achieved and the comparison pathway. It would include review effort and subsequent activity rather than counting only model inference or the number of suggested tests. The AMIE finding should therefore remain a rating result, not become an unsupported claim that doctors saved a particular amount of money.
The same caution applies when a plan suggests fewer actions. Lower activity could represent better prioritisation or inadequate care. Without understanding the clinical need and subsequent result, counting actions alone cannot identify value.
How to evaluate planning beyond a diagnosis list
A useful future assessment would follow the plan through implementation. Could the patient understand the next step? Was the proposed action completed? Did it resolve the relevant uncertainty? Were additional contacts needed because ownership or timing was unclear?
It should also examine what happens when the original plan becomes unsuitable. A patient may be unable to attend, prefer another option or develop new information that changes the clinical question. A practical system should support reassessment rather than treat deviation as a simple failure to comply.
These are proposed endpoints. They do not describe observed failures of AMIE, and they do not imply that manual clinical planning is consistently error-free. The comparator should be the actual existing service, with its own strengths and limitations.
What different readers should take from the finding
For clinicians, the result reinforces the value of checking the sequence and feasibility of a plan, not only whether its medical terms sound right. For researchers, it supports measuring implementation and patient consequences alongside rated reasoning. For learners, it offers a concrete exercise: turn a list of defensible options into a justified next step.
This comparison is published by iatroX and includes its own learning approach as an option for that educational task. As described in October 2026, the Socratic Tutor uses targeted questions after an attempted answer to explore the learner's reasoning. That design can support discussion of priorities, but it is not evidence that iatroX generates better clinical plans than AMIE or doctors.
Frequently asked questions
In which areas did clinicians' plans score better?
The March 2026 exploratory analysis, reiterated in the October researcher account, favoured clinicians on rated practicality and cost-effectiveness. Those results concern the assessed plans rather than measured service outcomes.
Did the study prove that AMIE ordered too many investigations?
No. The finding should not be converted into a claim about routinely excessive investigation or actual spending without specific supporting evidence.
What should a learner practise beyond generating a differential?
Practise prioritising the next decision, identifying missing information and explaining how the plan fits the patient and available pathway. A list of possibilities is only the starting point.