Dossier

Public AI — measure service value and preserve accountability

A framework for evaluating citizen-facing assistants and public-sector workflows.

Define the public outcome

Start with the service problem: unclear information, avoidable waiting, inaccessible processes or repeated administrative work. State whose experience should improve and how the change will be measured. The number of generated answers is an activity measure, not a complete public-value measure.

Separate information from decisions

An assistant that explains a procedure is different from a system that determines eligibility or changes a record. Make the authority of the output visible. Provide an accountable route to correct information and contest consequential actions where relevant.

Maintain the source of truth

Assign owners to the procedures and documents used by the service. Track effective dates and update paths. Test whether the system handles incomplete or conflicting information by asking for clarification or directing the user to the appropriate authority.

Evaluate inclusion

Test the service with people who use different devices, languages and accessibility tools. Examine failure rates and abandonment across user groups. Provide an alternative route when the digital interaction cannot complete the task.

Publish a meaningful service record

A useful public record explains purpose, owner, source material, evaluation approach, limitations and the route for feedback. Where publication is appropriate, include quality and service-completion measures alongside volume. Avoid publishing information that weakens security or exposes personal data.

Discussion for EventChanger

What evidence would demonstrate that an assistant helped a citizen complete a real task? Bring service owners, accessibility expertise and frontline staff into the same review, then publish a decision note with unresolved questions.