An NHS AI evaluation should follow people through access, assessment, decisions and outcomes, then examine who benefits and who encounters additional barriers. A favourable average cannot answer that question. The evaluation must include people who could not start, did not finish or required an alternative pathway, not only those whose interactions produced an analysable result.
Start with the inequality the workflow could create
Different tools create different opportunities for unequal effects. A patient-facing questionnaire might disadvantage people unable to describe their concern in its expected format. A scribe might preserve some speakers' clinical meaning less reliably. A prioritisation tool might inherit uneven recording of need. These are mechanisms to investigate, not claims that every deployment produces those effects.
The proposed evaluation should begin with a pathway discussion involving the people delivering and receiving the service. Identify where the tool changes access, information collection, human attention or the allocation of care. Then specify which groups and circumstances might be particularly affected at each point.
This avoids a generic demographic dashboard assembled after deployment. The dashboard should reflect a reasoned account of how the actual workflow could help or disadvantage people.
Count people before the first successful interaction
Define the eligible population independently of tool use. Distinguish being offered the service, attempting access, completing the interaction, receiving a decision and completing the next appropriate step.
A fictional digital assessment service might report excellent outcomes among completed forms while excluding people who abandoned an inaccessible question. Those people belong in the access analysis even when no clinical output exists to score. Their absence from the dataset is itself information about the pathway, not permission to ignore them.
HSSIB's July 2024 investigation into digital tools supporting patient access examined implementation and access concerns around online consultation. It was not a trial of modern generative AI, but it supports examining the whole service rather than assuming a digital route works equally well for everyone.
Choose groups for a reason, and protect the information
Relevant dimensions may include language needs, disability, age, deprivation, ethnicity, geography and the completeness of available records. Their relevance depends on the product and setting. Some analyses will require combinations of characteristics, while others will need attention to a specific communication or access barrier.
Do not collect sensitive information merely because it could populate another chart. Agree the purpose, appropriate governance, access controls and analysis plan before collection. Patients should not have to disclose unnecessary personal information just to obtain an alternative route to care.
The STANDING Together recommendations, checked on 10 October 2026, provide a useful foundation for documenting dataset representation and limitations. Their application should make uncertainty visible, not turn demographic categories into assumptions about an individual.
Use outcomes that can expose different kinds of harm
For each relevant group, consider access, completion, clinically important errors, unnecessary escalation, missed escalation, correction burden and the eventual resolution of the concern. Patient experience matters alongside clinical outcomes, but a pleasant interaction does not establish that the right action followed.
A proposed monitoring table might look like this:
| Pathway stage | Question to investigate | Response when a disparity appears |
|---|---|---|
| Entry | Who cannot use the offered route? | Review accessibility and the alternative pathway |
| Assessment | Whose information is lost or misinterpreted? | Inspect encounter examples and correction work |
| Decision | Who receives inappropriate or delayed action? | Reassess thresholds, context and supervision |
| Follow-up | Who remains unresolved after escalation? | Review ownership and practical access to care |
The table is an editorial design, not a national approval framework. Its purpose is to connect every measure to a possible decision rather than generate passive reporting.
Compare against real care, not a perfect baseline
An AI-assisted service may improve access for a group while still showing a gap compared with another group. Alternatively, an apparently equal distribution of outputs may conceal a worsening compared with the previous service. Both relative and absolute changes matter.
Use a comparator that reflects the actual alternative, including its delays, exclusions and errors. Where the deployment changes staffing or appointment availability at the same time, record those changes. Otherwise, improvements or harms may be attributed to the model when the surrounding service changed substantially.
The evaluation should also distinguish appropriate variation from inequity. Different clinical needs can justify different actions. The question is whether the difference is justified by relevant circumstances, not whether every group receives identical recommendations.
Small samples require caution, not silence
Publish subgroup denominators and uncertainty. Avoid ranking small groups by a single unstable percentage, and do not interpret the absence of a statistically significant difference as proof of equal performance.
A serious pattern in a small number of cases may warrant focused review even when the study cannot estimate its prevalence precisely. Qualitative accounts can identify a mechanism requiring investigation, but they should not be presented as population-wide error rates.
A reasonable proposed response is to combine quantitative monitoring with structured case review. Investigators can examine whether the same barrier recurs, whether it is plausibly linked to the workflow and whether a practical adjustment reduces it. Any remaining uncertainty should accompany the expansion decision.
Agree what will happen before results arrive
The team should specify who can restrict use, commission a targeted review or change the alternative pathway. A threshold without an owner is not an operational safeguard. Equally, a policy that waits for definitive proof of harm may be unable to respond to a credible emerging signal.
The WHO's January 2024 recommendations include post-release assessment and attention to disaggregated outcomes. These are recommendations to inform governance, not evidence that a local deployment has already met an equity standard.
Patient representatives should help interpret findings and proposed changes. The response may be better communication, redesigned intake, narrower scope or additional human support. It should not automatically mean withdrawing a useful service from the very group experiencing a barrier.
Keep learning improvements separate from outcome claims
A team might identify uncertainty about interpreting incomplete histories and organise a fictional case-review session. Source checking and guided questions can support that learning need. They do not, by themselves, demonstrate that the deployed system has become fairer.
The same distinction applies to iatroX. Its learning tools can provide a setting for examining assumptions and clinical reasoning, but any claim that they reduce inequalities would require a relevant evaluation. The credible sequence is to identify the barrier, change the workflow and measure what happens to the people affected.
Frequently asked questions
Is an overall accuracy score enough to assess health inequalities?
No. It can conceal differences in access, error severity and outcomes, particularly when people who cannot complete the pathway are excluded.
Does a non-significant subgroup difference prove the tool is fair?
No. A small sample may be unable to detect a meaningful difference, so sample size, uncertainty and plausible failure mechanisms still matter.
Should every disparity lead to immediate withdrawal?
Not necessarily. The response should reflect seriousness, evidence and available safeguards, with clear authority to restrict or stop use when continued deployment is not adequately controlled.