Before scaling a clinical AI pilot, an NHS organisation should establish that the actual workflow delivers a worthwhile benefit, with acceptable risks and manageable work, in the settings proposed for expansion. Usage and enthusiasm are useful signals, not sufficient evidence. The decision should connect a defined claim to a relevant comparator, representative evaluation and explicit conditions for continuing, restricting or stopping use.
Specify what scaling would change
Expansion can mean more users, a different specialty, a new patient population, additional permissions or less intensive supervision. Those are different changes. A successful pilot does not automatically support all of them.
Describe the proposed next stage before deciding whether the existing evidence is adequate. If the pilot involved a small, well-supported team and the expansion involves rotating staff across several sites, implementation and supervision assumptions need reassessment. If a draft-only function will begin taking actions, the task itself has changed.
The MHRA's intended-purpose guidance, published in March 2023, supports specifying the users, population and circumstances of use. A broad statement that a product is suitable for healthcare is not an adequate expansion specification.
Distinguish feasibility from effectiveness
A feasibility pilot can establish that the workflow can operate, identify implementation problems and estimate what a larger study would require. It may not be designed to establish improved patient outcomes or net cost savings.
Do not dismiss feasibility work for failing to answer a different question. Instead, decide what uncertainty remains before the organisation relies on the benefit at scale. A favourable experience report may justify a larger evaluation without justifying immediate organisation-wide adoption.
The DECIDE-AI reporting guideline, published in 2022, focuses on early live clinical evaluation of AI decision support and the human factors surrounding use. It helps make the pilot's contribution explicit, but it is not an approval pathway or a substitute for the appropriate regulatory and local assessments.
Compare the whole assisted workflow with realistic current care
An appropriate comparator includes the resources and support ordinarily available to staff. Comparing AI-assisted clinicians with clinicians denied their usual reference tools can answer an artificial question. Similarly, comparing generation speed with the time required to create and approve a final document can overstate benefit.
Measure the complete task: information gathering, checking, correction, completion and follow-up. Include work transferred to other staff and additional patient contacts. Distinguish increased access to previously unmet need from substitution for existing activity.
A fictional pilot might reduce time spent drafting letters while increasing the administrative effort needed to reconcile uncertain destinations. The overall result could still be worthwhile, but the business case should count both effects rather than report the visible saving alone.
Include the people who do not complete the ideal pathway
The evaluation should account for patients who decline the tool, cannot use it, are excluded, require escalation or disappear before the outcome is known. It should also include staff who use the system selectively or stop using it after initial training.
These groups are not automatically evidence of failure. Selective use may be appropriate, and some exclusions may be necessary. The question is whether the alternative pathway remains workable and whether reported results conceal the burden or outcomes of those outside the successful-use group.
The good machine learning practice principles, published in October 2021, emphasise representative evidence and relevant testing. Local expansion should make the limits of representativeness visible rather than assume that a larger licence count solves them.
Decide which outcomes would change the decision
A proposed scale-up scorecard should include the intended benefit, clinically important errors, correction effort, patient experience, equity, unresolved work and total resource use. Select measures that could lead the organisation to alter or stop the deployment.
Not every metric needs to improve. A system might increase staff review time while improving an important patient outcome. Another might save time while introducing an unacceptable failure. The decision should explain the trade-off rather than collapse all dimensions into a single favourable number.
Define critical safety signals and escalation routes in advance. Avoid both an impossible requirement for zero imperfections and a permissive rule that high adoption outweighs a serious unresolved risk.
The September 2026 policy proposals do not themselves authorise a deployment
On 10 September 2026, the National Commission into the Regulation of AI in Healthcare published recommendations including staged authorisation, continuing real-world monitoring and clearer public safety information. The announcement described recommendations for government and the MHRA to consider, not an already implemented universal approval route.
The policy direction is relevant to this discussion: evidence should develop as use expands, and monitoring should not end with initial access. However, a local organisation cannot treat the press release as permission to bypass its existing obligations.
This article's staged evaluation approach is an editorial proposal for decision-making as of 10 October 2026. It should be applied alongside current requirements and appropriate specialist advice, not represented as a newly enacted NHS standard.
Make expansion conditional and observable
A useful next-stage decision specifies the permitted sites, users and functions, the support required, the data to be collected and the review point. It should also state what would justify further expansion and what would trigger restriction.
Test whether the service can sustain the proposed oversight at the larger scale. A pilot supported by frequent expert attention may not predict performance when that support is diluted. The relevant question is whether the actual expanded arrangement remains workable, not whether the original demonstration was impressive.
Supplier changes also need a route back into the evaluation. If the product version or action scope changes materially, the organisation should reassess whether the pilot evidence still supports the claim.
Keep the final decision understandable
A board or partnership should receive a short account of the benefit demonstrated, the uncertainty remaining and the conditions attached to use. The underlying evidence can be detailed, but the decision itself should not require decoding a vendor presentation.
Where the pilot identifies a clinical learning need, source checking and targeted education can support staff development. iatroX can be considered within that separate educational role; it is not a substitute for the service's evaluation or a certification that another AI workflow is ready to scale.
The strongest expansion decision is not "the pilot was popular". It is "this defined use has earned this next stage, under these conditions, and we know what would make us reconsider".
What certification contributes to the decision to scale
A supplier's applicable certification and conformity documentation can be important evidence, but its contribution must be stated accurately. This section was added on 13 October 2026 and remains focused on the decision to expand an existing pilot, rather than initial supplier selection.
The European Commission's notified-body guidance describes assessment within defined scope. A scaling decision should establish whether the supplier's documents concern the product, functions and use proposed for the next stage. A management-system certificate should not be substituted for product conformity, and product conformity should not be treated as proof of the local benefit claimed.
For the fictional letter-drafting pilot discussed above, a relevant assessment may support the product's use under particular conditions. It does not resolve whether expansion will increase the administrative reconciliation burden or whether the new services can sustain the required review. Those questions remain part of the local evaluation.
A proposed supplier-to-adopter scale-up pack
The English NHS clinical-risk standards distinguish manufacturer and adopter responsibilities. A practical scale-up handover should therefore identify what can be reused and what changes at the next stage.
| Handover component | Contribution to the expansion decision |
|---|---|
| Product scope and applicable conformity documents | Establishes what assessment claims are relevant to the proposed use |
| Current configuration and change history | Links the pilot evidence to the release being expanded |
| Evaluation summary and unresolved questions | Shows demonstrated benefits and limits without hiding unfavourable findings |
| Supplier safety case and residual risks | Identifies assumptions and controls needing local ownership |
| Local workflow comparison | Makes differences in staff, setting, population and permissions explicit |
| Monitoring and response arrangements | Establishes who detects, investigates and acts after expansion |
| Training and fallback evidence | Shows whether the larger service can use and recover from the system appropriately |
This is an illustrative handover structure, not a complete official submission. Evidence should be proportionate to the proposed change, and existing reliable work should not be repeated merely to create a larger file.
An illustrative decision tree for the next stage
Start by asking whether the proposed scope is clear and supported by the relevant product and assessment documentation. If not, request clarification or pause the unsupported element before relying on the wider claim.
Next, ask whether the pilot evidence supports a worthwhile benefit in the proposed setting and whether consequential risks have adequate controls. Where both are sufficiently established and operational capacity is credible, proceeding within the defined scope may be reasonable through the organisation's applicable decision process.
Where evidence supports only a narrower population, function or level of supervision, consider proceeding with explicit restrictions rather than expanding the claim. The restrictions must be enforceable and monitored, not left as a qualification in a report that operational teams never see.
Where an important uncertainty can be addressed through further appropriate evaluation, request that evidence and define the decision it will inform. Where the risk cannot presently be controlled, or the service cannot sustain the required oversight, pause the proposed expansion and maintain a workable alternative.
These branches are editorial decision aids, not predetermined outcomes for an actual pilot. They complement the transparency principles, which support providing information relevant to limitations and local acceptance testing. As described on 13 October 2026, Insights by iatroX offers clinical safety and readiness support, but the adopter retains its own accountable decision about whether this evidence justifies this next stage.
Frequently asked questions
Does a successful pilot justify immediate organisation-wide adoption?
Not necessarily. Expansion may change the users, population, supervision or functions, so the evidence must support the proposed next stage.
Are the September 2026 commission proposals an implemented approval pathway?
The cited 10 September announcement presents recommendations for consideration, not an automatic authorisation for local deployment. Organisations must distinguish policy proposals from applicable requirements.
What should a scale-up decision contain?
It should identify the demonstrated benefit, realistic comparator, remaining risks, permitted scope, monitoring arrangements and criteria for further expansion or restriction. Responsibility for acting on the findings should be explicit.