An employer may introduce required workflows, but an instruction to use AI does not settle professional accountability, safety responsibilities or indemnity. Whether a particular requirement is lawful or reasonable depends on the circumstances. Clinicians need a clear account of the permitted task, available support and route for challenging unsafe use before routine deployment.
This article is dated 10 October 2026 and focuses on NHS workplaces in England, while recognising that GMC standards apply across the UK. It is not advice on an individual employment dispute. The GMC's current AI guidance specifically addresses concerns about being asked to use technology believed to be unsafe.
Mandatory adoption is not a complete operating procedure
A fictional department announces that every outpatient consultation should use its new AI scribe. Staff receive login details and a short demonstration. Nobody explains whether use is expected when a patient objects, an interpreter is present or the clinician cannot complete review before the next appointment.
The problem is not necessarily the decision to standardise a useful tool. It is that the instruction leaves consequential decisions unresolved. "Use the scribe" does not specify what recording is authorised, what output may enter the record, who checks it or what to do when the workflow becomes unsuitable.
A clear implementation should describe exceptions as well as normal use. Otherwise clinicians may interpret compliance as using the tool regardless of context, or avoid it entirely because they cannot establish the boundary. Neither response demonstrates a well-designed clinical process.
Allocate tasks that people can actually control
Responsibility should be attached to an activity and the ability to influence it. A clinician can assess a clinical recommendation and review a visible draft. They may not be able to inspect a hidden integration, change retention settings or prevent an unannounced supplier update.
That does not mean the clinician has no duty to report a concern in those areas. It means the organisation must identify who can investigate and change the relevant system. An accountability chart that assigns everything to the final signatory is not a substitute for operational ownership.
A proposed allocation exercise should trace the workflow from procurement to patient communication. Identify the product owner, clinical safety lead, information-governance route, technical support, responsible reviewer and person authorised to restrict use. Where one person holds several roles, confirm that they have the competence, authority and time required.
Training should establish capability, not attendance
A demonstration shows how a product is intended to work. It does not establish that a clinician can recognise an omission, preserve uncertainty, handle patient dissent or recover from a failed transfer.
For a fictional training exercise, give the learner a synthetic consultation and a plausible draft containing an attribution error. Ask them to explain what should change and whether the result is ready for the record. Then introduce an unavailable source or an uncertain patient identity. The aim is to test the actual review task, not proficiency in clicking through the interface.
Include temporary staff and people returning after an absence. Access to a previous training session does not establish familiarity with the current configuration. Equally, avoid treating confidence with software as evidence of clinical competence or treating a request for training as resistance to innovation.
Review time is a resource decision
An organisation cannot meaningfully require verification while scheduling work on the assumption that no verification is needed. In the fictional department, clinicians may discover that complex consultations require more checking than the demonstration suggested. That finding should influence workflow design, not be dismissed as slow adoption.
Measure the actual work: reading, comparison with the encounter, correction, approval and management of uncertain content. Consider whether the task is easier immediately after the consultation or whether the intended schedule creates delayed review with poorer recall.
GMC record-keeping standards, checked on 10 October 2026, require clear, accurate and contemporaneous records, with appropriate detail. They do not turn an unrealistic appointment template into a safe review process. The organisation should create conditions in which the required standard is achievable.
Questions to settle before required use
The following is a proposed discussion framework for clinical teams, managers and staff representatives. It is not an exhaustive legal checklist.
| Topic | Question requiring an operational answer |
|---|---|
| Scope | Which functions are required, optional or prohibited? |
| Exceptions | What happens when the patient or consultation does not fit? |
| Review | Who checks the complete output, and when? |
| Competence | What training and reassessment are available? |
| Technical defects | Who can investigate, restrict or roll back the system? |
| Indemnity and support | Which arrangements cover the actual work and related proceedings? |
| Reporting | How can staff raise concerns without the issue being reduced to a usage target? |
Written answers reduce ambiguity, but should not create false reassurance. A policy can be clear and still describe an inadequate process. Test the workflow with the people expected to use it, including those who receive its outputs downstream.
Challenge unsafe use through the right route
GMC guidance on raising concerns supports prompt action when systems or policies may put patients at risk. The immediate response should depend on the seriousness and urgency of the concern. A clinician should not be left choosing between silent compliance and an improvised confrontation during a consultation.
A practical report describes the observed issue, the affected task, why the existing control may be insufficient and what temporary action is proposed. Separate direct observation from inference. For example, "The review screen omits the attachment" is more actionable than "The AI is unsafe".
Where an employment or contractual dispute develops, seek appropriate professional, representative or legal advice. This article does not establish an automatic right to refuse every AI-related instruction, nor suggest that an employer's instruction overrides professional duties.
Keep learning separate from performance surveillance
A team may use fictional cases and source checking to improve its understanding of the new workflow. That is different from inferring competence from logins, prompt volume or willingness to accept a draft.
This distinction applies to iatroX as a learning platform too. A clinician can use a general reference question or a reflective activity to explore uncertainty, but activity data alone do not prove readiness to supervise another product. An organisation should not substitute a completed learning exercise for task-specific training and adequate support.
The constructive outcome is a shared operating agreement: staff know what they are expected to do, the organisation supplies the conditions to do it properly, and both retain a credible route for changing the process when it does not fit the patient or the task.
Frequently asked questions
Can a doctor refuse every employer instruction involving AI?
This article cannot establish a universal right of refusal; the facts, contract, law and safety concern matter. Raise specific concerns through appropriate routes and obtain advice where necessary.
Does employer approval remove a clinician's professional responsibilities?
No. It also does not remove the organisation's responsibilities for the deployment and activities within its control.
What should be agreed before mandatory use starts?
At minimum, clarify scope, exceptions, training, review arrangements, technical escalation and relevant indemnity or support. Those agreements should describe the real workflow, not merely confirm access to the software.
Build a focused learning activity around an AI supervision question →