Assess what happens to customers excluded from AI
An AI safeguard can reduce problematic answers while leaving redirected customers without useful support. Assess the alternative route and retain excluded customers in the service outcome record before expanding automation.
Keep sources linked to the record used for review. The diagram does not represent automatic approval.
A customer excluded from an AI journey still has a service need. If the assessment ends when the model declines or redirects, the organisation cannot tell whether its safeguard prevented harm or merely moved the unresolved work elsewhere.
The Financial Conduct Authority (FCA) published its AI practice review on 19 October, with a page update dated 20 October. It warns that automatically excluding consumers with vulnerability indicators can deny appropriate support or benefits when a suitable alternative is absent. Its examples describe developing practice under existing rules and create no new requirements.
For a customer-service owner, the useful decision is whether the combined AI and alternative routes improve the service. Keep customers affected by exclusion in that assessment. This does not mean admitting everyone to automation. It means establishing what the organisation delivers when automation is unsuitable.
Follow the need beyond the exclusion
Consider a hypothetical financial-services support tool that helps customers understand routine account correspondence. It redirects anyone who mentions difficulty using a telephone to a telephone support line, because its screening rule treats that statement as a reason to stop the AI interaction.
The guardrail fires exactly as configured. The destination is nevertheless unusable for the stated need. A dashboard counting successful redirections could report a favourable control result while the customer remains unable to obtain help.
A more useful design would preserve the reason support is needed, identify an appropriate available channel and tell the customer what will happen next. Where the service offers secure messaging or a staffed written route, test whether the customer can actually reach it. Where no suitable route exists, record the service gap for its owner rather than marking the case complete.
The opportunity is a more flexible support service. AI could remain available for bounded explanation where suitable, while people receive a usable alternative when the task or their preferences call for it. An exclusion rule should be specific enough to explain which activity it restricts and why. A restriction on automated recommendations need not automatically remove access to approved factual information.
The FCA’s older guidance summary on vulnerable customers describes vulnerability as a spectrum and recognises that needs differ. Its July update also directs firms to current Consumer Duty material. That supports examining the particular support need rather than treating an indicator as a complete description of the person.
Keep both routes in the outcome record
For a proposed assessment, define the intake before the exclusion rule runs. Retain the original service request and record whether it continued through AI, moved to another channel, was abandoned or remained unresolved. A customer who leaves before receiving support must not disappear from the denominator merely because the AI system stopped responding.
Report model behaviour separately from the service outcome. Correctly applying a restriction, reaching an adviser and resolving the customer’s need are different events. Keep the timing and evidence for each. If the eventual outcome cannot be observed, report it as unknown and investigate the missing link.
| Proposed observation | Decision it supports |
|---|---|
| Reason for restriction and requested support | Whether the rule addresses the actual task and need |
| Alternative offered and whether it was reached | Whether redirection creates a usable route |
| Resolution, abandonment or unknown outcome | Whether support continued after the AI interaction |
| Repeated explanation and time to useful help | Whether the channel change added avoidable burden |
These are proposed service measures, not a prescribed FCA reporting template. Use a defined observation period and retain relevant differences between customer needs and request types. Compare like requests where possible. A group receiving more complex cases may take longer without the channel itself causing the delay.
Do not turn this exercise into indiscriminate collection of sensitive information. Agree the minimum case data, access and retention with the relevant owners. Reviewers need enough context to understand the route, without copying an entire conversation into every downstream system.
Decide on the whole support route
Swipe or scroll for the full diagram →
What happens after the AI journey excludes a customer?
Owner: Customer-service owner. Threshold: The alternative is reachable, appropriate to the customer's need and followed through to a recorded outcome.
- Expand
- Retain the bounded AI route where evidence supports it and the alternative continues to meet the needs of redirected customers.
- Retain narrower scope
- Repair or replace an inaccessible alternative before expanding automation or reducing existing support capacity.
- Gather more evidence
- Investigate abandoned, unresolved and unobserved cases. An exclusion count alone cannot show that support was delivered.
A proposed decision rule for a hypothetical support service. It is not an FCA requirement, an implemented control or a measured customer outcome.
A fictional support service redirects a customer away from AI. Continue only if an appropriate alternative can meet the customer's need. Repair the alternative when access fails and investigate missing outcomes before claiming the safeguard improved service.
Test the alternative with the people who need it
Start with constructed cases that exercise the proposed restriction and the alternative. Include a customer who cannot use the suggested channel, someone who explicitly asks for a person and a request whose urgency makes a delayed response unsuitable. Have frontline staff and relevant accessibility reviewers assess whether each route can meet the stated need.
Then observe an appropriately scoped service comparison. Preserve existing support while evidence develops. Review excluded and abandoned cases alongside completed AI interactions, and test whether information already provided survives an authorised channel change accurately. A generated summary should remain distinguishable from the customer’s original words.
The FCA review’s consumer-support section describes continuity between channels as good practice and identifies inflexible journeys as a problem. It also cautions against assuming a language model reliably recognises vulnerability from text. A missed indicator therefore belongs in the assessment alongside an unnecessary exclusion.
The archive’s live-testing article establishes a controlled boundary for exposing customers to AI. This assessment follows the customers left outside that boundary. It asks whether the protection creates a workable service path, including where the model is deliberately prevented from answering.
Make expansion depend on the alternative
The practice review does not quantify how common these failures are or establish the effect of a particular routing design. The example and assessment method here are proposals, not observed Institute results. The firm’s compliance team and qualified advisers should determine the applicable obligations for its service.
The customer-service owner should approve the exclusion rule together with the alternative, its capacity and the evidence needed to judge outcomes. If the AI route improves while redirected customers lose access or remain unobserved, repair that gap before reducing existing support or expanding the deployment. A smaller automated scope can be a sound decision when the whole service still meets the needs of the people it serves.