What happens when someone cannot use your AI service?
An AI-enabled public service is not inclusive because a human channel technically exists. People must be able to find, reach and complete the service through it without penalty.
In today’s governance note, I want you to consider what happens when someone cannot, or reasonably does not want to, use the AI-enabled route through your service. They may have a disability that makes the interface difficult to navigate. They may lack reliable internet access, digital confidence or the language skills the system expects. They may be trying to explain circumstances that do not fit neatly into a chatbot exchange. They may simply be unwilling to share sensitive information with an automated system.
Can they still complete the service? This question should arise before the AI route is selected, while the organisation defines the service model, budget, and staffing requirements. If it appears only during accessibility testing, the alternative is likely to be treated as an exception rather than part of the service itself. A telephone number at the bottom of a page is not necessarily an accessible alternative. Nor is it a human channel that requires longer waiting, repeated explanations or additional evidence. If avoiding AI makes the service more difficult, more expensive or less likely to produce a timely outcome, the person does not have a meaningful choice.
I would begin by mapping the whole journey rather than the interface. Identify where AI first enters the process, what information it collects, which steps it controls and where its outputs influence what happens next. At each point, decide how someone can move to another route without losing the information they have already provided or being returned to the beginning. The alternative may involve telephone, written or in-person support. It may involve a staff member completing the digital process with the user. The appropriate arrangement will depend on the service, but it must be designed, funded and tested alongside the AI route. You also need to make the alternative visible.
Do not require someone to fail repeatedly before presenting human support. Explain clearly when AI is being used, what it can help with and how to reach another channel. The person should not have to successfully use the system to discover how to avoid it. Measure the experience across every route. An AI channel may appear successful if the organisation tracks only the number of conversations it contains, without transferring them to staff. That figure tells you little about people who abandon the service, repeat the task elsewhere or accept an unsuitable response because the escalation route is unclear.
Useful measures include completion, abandonment, repeat contact, waiting time, successful transfer, complaints and whether outcomes differ by channel. Examine why people leave the AI route and whether the service itself is creating the need for additional support.
The UK Government Service Standard requires public services to work for disabled people and for those who lack internet access, skills or confidence. Its assisted-digital guidance, updated on 18 December 2025, states that support needs should be researched and incorporated into the service. It also requires teams claiming that no users need support to demonstrate this through research and testing. GOV.UK Service Standard, assisted-digital guidance. The guidance distinguishes accessibility from assisted digital support. An interface can meet technical accessibility requirements while remaining unusable for someone who lacks confidence, connectivity or the ability to express a complex situation through it. Institutions need to address both.
Offering another route recognises that public services carry obligations that ordinary products do not and ensures everyone who needs the underlying service can still obtain it.