The fallback must be a person or team with the knowledge, information, time and authority to resume the work. Naming a clinician as the supervisor does not establish any of those conditions. Nor should an organisation assume that AI has caused deskilling without evidence: retained competence and readiness to take over are separate questions that both need assessment.
A capable doctor can still be unable to take over
In a fictional outpatient service, an assistant normally gathers information, prepares a summary and organises the next administrative steps. During a clinic, the assistant becomes unavailable. The supervising doctor remains clinically knowledgeable, but cannot see which patients completed intake, which questions remain unanswered or which messages have already been sent.
This is not necessarily deskilling. It is a loss of situational understanding and access to the state of the workflow. Asking the doctor to "revert to manual working" provides no answer to whether a patient has already received instructions or whether an action is still pending.
A second problem could coexist: some staff may rarely have performed the manual process. That may reflect incomplete initial training rather than deterioration of an established skill. A third is capacity: the team might know how to perform the work but lack enough available staff to handle the volume without assistance.
A useful fallback assessment separates these problems instead of giving all of them the same label.
What the deskilling evidence does and does not show
A multicentre observational colonoscopy study published online on 12 August 2025 compared unaided performance before and after the introduction of AI assistance at participating centres. It found a reduction in adenoma detection during non-AI procedures after exposure, raising concern about changes in endoscopist behaviour.
The study was retrospective and task-specific. It does not establish universal, irreversible loss of clinical skill, or show that language-model agents cause the same effect in general practice. Its relevance is to the need to examine retained unaided performance rather than assuming that assisted performance answers that question.
For a service introducing AI, the corresponding evaluation should specify the competence that matters. Is it interpreting a case independently, recognising an inappropriate recommendation, operating a manual process or recovering a partly completed task? These require different exercises and may involve different professional groups.
Build a fallback around the work, not the off-switch
A proposed fallback plan starts with the essential service functions. Which activities can continue using approved alternatives? Which require a different mode of assessment? Which can safely wait, and which need immediate reassignment? These decisions should be made by the appropriate clinical and operational leads before an outage.
The 2025 SAFER Guides include contingency planning for partial or complete electronic-record unavailability. They provide a useful US health-IT safety framework, not a UK approval scheme or a certification that a particular AI fallback is adequate.
For the fictional clinic, staff should be able to identify the current queue, the latest confirmed patient information and actions already completed. A locally agreed handover record might state the original task, its status, the evidence of completion, remaining uncertainty and the person now responsible.
The alternative should not depend entirely on the same failed component. Keeping instructions only inside the unavailable assistant would undermine their purpose. Equally, creating uncontrolled copies of patient data in personal accounts would introduce a different risk rather than solve the first one.
Rehearse partial failure as well as total failure
A full outage may be easier to notice than a degraded service. A tool could still generate prose while its record connection is unavailable, or display an old summary without making its age obvious. Staff need a clear way to know which functions remain supported.
A proposed tabletop exercise can present a partial failure without involving real patient care. The team is told that the assistant can read previously loaded information but cannot confirm whether new actions succeeded. Participants must decide what to pause, what to verify elsewhere and how to avoid repeating an uncertain action.
Another exercise can remove the original user. A colleague covering the clinic should be able to understand the task without relying on that person's private recollection. A third can introduce changed patient circumstances, testing whether staff recognise that completing the old plan is no longer the immediate goal.
The exercise should assess correctness and recovery burden, not just speed. Restarting quickly is not success if it creates duplicate communications or leaves an unresolved clinical issue hidden.
Protect the capabilities the service still depends on
Where clinicians are expected to judge an AI's interpretation, education should retain opportunities to formulate and explain an interpretation independently. Where they must examine patients, practical supervised experience remains necessary. Where administrative staff provide operational fallback, they need access and training rather than a vague expectation that they will improvise.
This does not require manually repeating every automated task in routine care. A proportionate approach could use periodic simulation, supervised case discussion and selected independent practice outside the live workflow. The frequency and content should reflect the consequences of failure and evidence about the team's needs.
Assessment should distinguish a temporary unfamiliarity with a screen from a clinical knowledge gap. It should also identify organisational barriers. A clinician who cannot obtain the original evidence because access was never provided should not be treated as having failed a competence test.
Any proposal to test retained ability should protect patients. Necessary support should not be withdrawn from real care simply to create an experimental comparison.
Plan for a smaller service during disruption
Sometimes the safe fallback is not the same throughput without the software. The organisation may need to reduce activity, redistribute work or use another service while restoring reliable operation. Pretending that existing staffing can absorb every automated task can turn a technical failure into unmanageable workload.
The plan should state who can make that decision and how patients and partner services will be informed. It should also cover recovery: reconciling work performed during downtime with the restored system and confirming that queued actions will not execute unexpectedly.
These are proposed operational safeguards. Their adequacy must be tested in the actual service, including the arrangements for locums, out-of-hours teams and staff with different access permissions.
Practise judgement, without claiming a software antidote
As described by iatroX in October 2026, its question banks, Socratic Tutor and simulations offer different formats for practising knowledge application and explaining decisions. They can supplement a proposed learning programme for clinicians who supervise AI-supported work.
They do not establish that a clinician is competent to operate another product, prevent deskilling or replace bedside and procedural practice. A useful learning record should identify the skill practised, what changed and what still needs supervised assessment. The strongest human fallback is not confidence in the clinician's job title; it is demonstrated capability within a workable service plan.
Frequently asked questions
Does AI inevitably deskill clinicians?
No. Existing evidence concerns particular tasks and settings, and does not establish universal or permanent deterioration in clinical ability.
Is manual fallback mainly a training issue?
Training is one component, alongside access to current information, operational capacity and authority to change the workflow. A trained clinician may still be unable to recover safely if those conditions are absent.
Can simulation replace the practical experience needed for fallback?
Simulation can rehearse reasoning, handover and recovery decisions, but it does not replace the supervised bedside or procedural experience required for those clinical tasks.
Practise explaining a decision before relying on assistance →