Find the Customers Your AI Pilot Leaves Out

Ask the service owner, customer-research lead and risk team to map the route from first contact to a useful result. Include discovery, registration, identity checks, task completion, human help and appeal. Record the denominator at each stage. A model’s answer accuracy belongs inside that journey; it cannot stand in for the whole journey.

Choose a small number of groups whose experience could plausibly differ, based on the service and existing evidence. Recruit people who dropped out, not only enthusiastic pilot users. Collect sensitive information only where there is a lawful, necessary and proportionate basis. Missing data should remain a limitation, not become a reason to infer a person’s identity.

World Teachers’ Day also falls on 5 October. UNESCO’s 2026 programme emphasises teachers’ professional status, autonomy and voice. For education providers, that is a separate reason to involve teachers in reviewing AI-supported services. It does not establish that a particular teaching tool works. ²

Bring one concrete failure to the review: an inaccessible verification step, a misunderstood explanation or a hand-off that ends without help. Ask who can change it, what it costs and how the customer will know it has improved. A workshop without authority to change the service is unlikely to answer the rollout question.

CGAP’s August account of work with financial institutions in Mexico and Morocco makes a useful distinction between institutional preparation, operational change and customer outcomes. Teams and action plans matter, but they are not proof that women are better served. This is early implementation experience in two countries, not an AI trial or a global impact estimate. ³

For organisations involved in Northern Ireland’s public sector, there is also a real deadline: consultation on its draft AI strategy closes on 7 October at the listed time of 5pm. Relevant teams should decide whether they have evidence to submit. This is a consultation, not enacted law or a deadline applying to other markets. ⁴

Mexico and Morocco offer useful institutional experience, but neither represents an entire region. Similarly, a well-performing English-language service in North America or the UK cannot establish performance in another language or financial system. Local partners should help define the test, not merely translate the interface.

For operations across Asia, Africa, the Middle East and Latin America, use this week’s regional discussions to identify questions about documents, connectivity, language and access to human support. These are hypotheses to investigate in each service population, not assumptions about everyone in a region. Compare observed barriers before choosing a technical fix.

Keep an alternative route available while investigating. It may involve a person, a simpler digital journey or an existing service channel. Measure its cost and quality too. Inclusion does not require every customer to use AI, and a non-AI route should not quietly become an inferior service.

Return to the rollout decision with three findings: who was missing, which barrier was changed and what happened when affected customers tried again. Report unresolved cases alongside successful completions. Keep the sample size, recruitment method and observation period visible; a small discovery exercise identifies problems rather than proving population-wide impact.

Expand only the part of the service for which the evidence is adequate. If a group still cannot complete the journey or obtain help, assign the fix and retain the alternative route. If the suspected disparity disappears under a fair comparison, revise the concern rather than preserving it to justify the project.

By Friday, the useful output is a narrower, better-supported operating decision—not a declaration that the organisation has solved inclusion. The strongest AI pilot may be the one that reveals whom its own success rate has left out.