A responsible clinician or team still needs to own the result and its follow-up; an AI inbox label is not a transfer of clinical responsibility. The exact allocation depends on the service and agreed handover arrangements. The practical test is whether someone can identify what remains unresolved, who must act and what happens if that person is unavailable.
Start with the result pathway, not the inbox screen
For UK practice, the BMA's guidance on acting upon electronic test results, reviewed on 10 October 2026, advises that the ordering clinician should generally be assumed responsible unless responsibility is appropriately transferred. It warns against assuming that another clinician will act simply because they can view a result. It also recognises that receiving or seeing an urgent result can create a need to act or ensure action.
These principles require explicit local arrangements. They do not justify ignoring a result because somebody else ordered it, or automatically transferring every hospital result to a GP because an electronic copy arrived.
AI changes the mechanics of sorting, summarising and routing, but not the need to resolve those relationships. An agent's assignment of a task to a named inbox should be treated as an operational event whose meaning must be defined, not as a legal conclusion that responsibility has safely moved.
A fictional amended-report problem
Imagine a hospital report initially marked as requiring no additional action. An amended report subsequently identifies a finding that needs review. A hypothetical inbox assistant groups the second message with the first and preserves the earlier "reviewed" status.
The original clinician may have acted appropriately on the information then available. The new problem is that a changed report has inherited a completed state. If the system counts messages organised rather than clinically distinct results reviewed, the amendment could disappear from the active workload.
A proposed safeguard is to treat a consequential amendment as a new review event linked to the earlier result. The reviewer needs to see what changed, which previous action depended on the original report and whether anyone must now be contacted. This is a constructed failure scenario, not a report about a named supplier.
Separate the milestones that an agent might collapse
A result can be received, recognised, routed, acknowledged, clinically reviewed, communicated and acted upon. Those are different states. Acknowledging a task may establish that someone has seen it, but not that they have interpreted the result or completed the plan.
Similarly, "patient informed" is incomplete if it means only that a message was generated. The service needs an appropriate account of what was sent, whether the channel was suitable, whether further contact was required and what the patient was expected to do.
The HSIB investigation published in 2019 into unexpected significant radiological findings examined failures of communication and follow-up before current agentic systems. It establishes that this is an existing patient-safety problem, not one invented by AI. Automation should be compared with the real pathway and its known weaknesses, not an imaginary error-free manual inbox.
Design an unresolved-results queue
A proposed operational queue should show the result, the reason it remains open, its responsible team, the next required action and the clinically appropriate escalation arrangement. It should distinguish waiting for clinician review from waiting for patient contact, further investigation or information from another service.
A single overdue flag is insufficient if all pending work is treated identically. A clinically urgent unreviewed result and an agreed routine follow-up have different implications. Time limits should be locally defined for the task and clinical situation, rather than copied from a generic AI configuration.
Staff absence also needs an explicit route. Work should not remain assigned to a person who has left the organisation or is unexpectedly unavailable. Covering staff need enough context and authority to act, not merely permission to view the original clinician's name.
These are proposed design requirements, not claims that one particular implementation already provides them. A buyer should ask for a demonstration of the exception pathways as well as the routine sorting function.
Test transfer across organisational boundaries
A handover should identify the required action, the relevant clinical context and the receiving service. Where a workflow depends on another team accepting responsibility, the sending organisation needs a way to establish the handover's status and respond if it fails.
This is more specific than an email appearing in a sent folder. A destination might be technically valid but operationally inappropriate, or a message might reach a general mailbox without entering the relevant clinical process.
For the amended-report scenario, the test should include a patient who has already been discharged and an ordering clinician who is no longer on duty. Who reviews the change? Who contacts the patient? Who establishes whether the previous plan needs revision? The system should expose those questions rather than supply an unsupported "complete" label.
The GMC's delegation and referral guidance, reviewed on 10 October 2026, provides a professional framework for clear communication when work passes between clinicians. An AI routing function should support those arrangements, not be mistaken for a professional accepting the delegation.
Include non-response and missing results
A patient who does not reply is not necessarily a patient who has understood and declined the plan. The proposed workflow should allow staff to consider communication barriers, incorrect contact details and the clinical significance of non-response. It should not automatically repeat the same ineffective message indefinitely.
There is a separate question about results that never arrive. An inbox can process every incoming item perfectly while an expected investigation remains outstanding elsewhere. A service-level assessment should therefore include reconciliation of expected and received results where appropriate, rather than measure only what entered the agent's view.
These distinctions also affect reporting. A high proportion of inbox items closed cannot establish that follow-up is complete if missing results, unsuccessful contacts or amended reports are excluded from the denominator.
Use learning tools for the clinical question, not task ownership
A result-management review may identify a knowledge question that warrants further study. As described by iatroX in October 2026, Ask-iatroX supports general clinical-reference enquiries with linked sources. It is not presented here as an inbox integration, a result-tracking service or a mechanism for transferring responsibility.
A fictional teaching exercise can ask learners to identify what is missing from an apparently completed pathway: interpretation, communication, an accepted handover or confirmation of follow-up. That develops a useful distinction without entering identifiable patient records into a separate learning environment.
The final question for procurement is concrete: can the service account for each unresolved result, including during absence, disagreement and technical failure? Faster sorting matters only within a pathway that can answer it.
Frequently asked questions
Does assigning a result to an AI-managed inbox transfer responsibility?
Not by itself. Responsibility must be clear within the clinical service and any agreed handover, with appropriate arrangements for review and follow-up.
Is sending the patient a message enough to close the result?
It depends on what the clinical plan requires and whether the communication is appropriate and sufficient. Message generation or delivery alone does not establish completed care.
Should a clinician ignore an urgent result ordered by another service?
No. The BMA guidance recognises that a clinician receiving or viewing a result requiring urgent action may need to act or ensure that it is actioned, even where another clinician ordered it.
Explore the clinical reasoning behind a general results question →