Healthcare agents should be judged partly by when they hand back, not only by how often they finish without help. Selective autonomy means allowing a system to handle a defined subset while referring other cases for review. Its value depends on both the decisions it retains and the quality of the pathway receiving everything else.
A Nature Medicine study published on 15 September 2026 provides a concrete research example. In a retrospective diagnostic simulation, one reported threshold retained 49.4% of cases at 98.9% diagnostic accuracy. Both figures are necessary: the accuracy describes the retained subset, not every presentation.
Accuracy without coverage hides the decision
A system can improve its reported accuracy by withholding difficult cases. That may be a sensible design choice, but the result cannot be interpreted without knowing how much work is retained and which cases are deferred. A high score on a selected group is a different claim from high performance across the whole population.
Coverage is the proportion of cases the policy allows through without the specified review. Accuracy describes agreement with the reference outcome within the evaluated group. Neither figure alone establishes the safety of the whole clinical service, because the deferred group still needs appropriate care.
The September 2026 study's 49.4% and 98.9% result therefore does not mean that roughly half of real primary-care work can safely be automated. The setting was simulated diagnostic reasoning with particular cases and reference labels. The clinical work and operating conditions of a live service are broader.
A reliability signal is not permission to practise
The study investigated signals associated with whether a decision was correct, including consistency across repeated runs. Its promising finding concerns discrimination within that evaluation, not a universal threshold for all clinical systems. The authors identify the need for prospective evaluation and configuration-specific calibration.
A reliability signal can inform a proposed routing policy. It does not grant clinical authority, establish eligibility or replace the rules governing what the product may do. A highly consistent answer can still concern an action outside the system's permitted scope.
The FDA's January 2026 clinical decision-support guidance reinforces the importance of considering the actual software function. In practical terms, a decision about whether a model seems reliable should remain separate from whether the service is authorised and equipped to act on that decision.
The deferred cases do not disappear
A selective system creates a second workflow: human review. That workflow needs available staff, relevant information, clear ownership and a response appropriate to the patient's situation. An agent that refuses confidently but hands back poorly has not necessarily made the overall service safer.
Consider a fictional clinic where the automated pathway accepts straightforward requests but sends ambiguous cases to an unprioritised inbox. The retained group may perform well while the deferred group waits without clear responsibility. A headline about autonomous accuracy would miss that service-level problem.
A better hand-back would explain the original goal, relevant evidence, missing information, actions already taken and the decision still required. The reviewer should be able to continue without reconstructing the entire interaction or unknowingly repeating an external action.
Four outcomes belong on the same scorecard
A proposed scorecard should include inappropriate autonomous handling, appropriate autonomous handling, justified escalation and avoidable escalation. These categories preserve the distinction between an error that remains with the agent and work that is unnecessarily transferred to a person.
The consequences also matter. A minor administrative correction and a clinically consequential wrong decision should not be collapsed into one unweighted error count. A service should explain how severity was assessed and whether the denominator includes incomplete or interrupted cases.
For escalations, measure whether review was actually completed and what happened next. The GMC's delegation and referral guidance, checked on 9 October 2026, provides a useful UK professional anchor for clear transfer of work. It does not validate a particular model's hand-back policy.
Do not assume human review is perfect
A threshold policy often implies that difficult cases will become safe once a clinician sees them. That assumption deserves testing. Reviewers can face incomplete records, time pressure or persuasive but unsupported explanations. The quality of the hand-back can influence how effectively they respond.
A fair comparison should therefore examine the clinician-plus-system workflow, not combine measured model performance with an imagined error-free reviewer. It should also compare against the actual alternative, including normal reference access and existing triage arrangements.
This does not make human review pointless. It means review needs to be designed as clinical work rather than treated as a reassuring label. The objective is an effective partnership with understandable boundaries, not an apparently autonomous system whose difficult work becomes invisible elsewhere.
Who is selected out matters
An acceptable overall coverage figure can conceal uneven access. Some groups may be deferred more often because their records are less complete, their language differs from the tested examples or their needs are more complex. Those are potential mechanisms to investigate, not conclusions about a particular product.
A proposed evaluation should report retained and deferred performance across relevant patient groups, alongside waiting times and outcomes after review. Equal autonomous coverage is not automatically the correct goal, but unexplained differences should not be hidden inside an aggregate.
The question is whether the routing policy gives each patient an appropriate pathway. A lower autonomous rate could represent sensible caution or an avoidable access barrier. Distinguishing the two requires clinical context and information about the human service that receives the case.
Test the policy before treating it as a stable property
A practical evaluation can compare several candidate thresholds on a held-out dataset, then test the selected policy in a governed setting without assuming that retrospective performance will persist. The protocol should specify how changes in model version, input quality or tools trigger reassessment.
The stopping rules should also be explicit. Repeated consequential errors, unmanageable review queues or systematic failure in a subgroup could require narrowing or pausing the automated pathway. This is an evaluation proposal, not a report that the September study deployed such a service.
The useful output is an operating policy with evidence and limits, not a single reassuring confidence number. It should say which cases can proceed, which must return to a person and what the service does when those assumptions no longer hold.
What clinicians can practise
A learning exercise can ask a clinician to identify what would make them retain, challenge or return a proposed action. The important skill is explaining the boundary, including what information would change the decision.
Under iatroX's September 2026 specification, Tutor supports targeted reasoning around an attempted question. That can contribute to deliberate practice, but it is not certification to supervise an autonomous service and has not been shown to make a particular threshold safe.
Selective autonomy is therefore a constructive alternative to all-or-nothing automation. Its success should be judged by the complete routing system, including the cases it does not handle, rather than the impressive accuracy of the subset it keeps.
Frequently asked questions
What does selective autonomy mean in healthcare?
It means permitting a system to handle a defined subset while routing other cases for review. The selection policy and the receiving clinical workflow both need evaluation.
Does 98.9% diagnostic accuracy apply to every case in the cited study?
No: the September 2026 study reports that figure for a retained subset covering 49.4% of cases at one threshold. It is a simulated diagnostic result, not whole-service clinical accuracy.
Is handing back always the safest action?
Not automatically: unnecessary or poorly managed escalation can delay care and create work. The hand-back must be justified, informative and connected to an effective response.