Designing AI-Enabled Services for Citizens

Public sector organizations increasingly deploy conversational assistance on benefits applications, licensing systems, and permit services. These deployments frequently report strong performance against interaction measures: questions answered, sessions handled, satisfaction scores recorded. Completion rates in many cases remain unchanged.

The mechanism is straightforward. A conversational assistant answers questions about a process. It does not alter the process. Where an applicant abandons a submission at the fourth step because a required document was never identified at the first, accurate answers earlier in the sequence do not address the cause of abandonment.

This distinction determines how public facing artificial intelligence should be evaluated.

Completion is the measure the agency shares with the public

Agencies track page views, session duration, conversation volume, and satisfaction. These measure the service’s activity rather than the public’s success.

Completion rate measures the proportion of individuals who begin a task and finish it. It is the single measure on which the agency’s interest and the applicant’s interest are identical.

Measuring it requires step level instrumentation rather than an aggregate abandonment figure: which field, which document requirement, which eligibility question. In most services, abandonment concentrates at three or four specific points.

Those points are frequently identified by frontline staff before analytics surface them. Consultation with contact center personnel should therefore precede formal research, as staff handling exceptions can generally identify the difficult steps directly.

Establish the abandonment cause before selecting capability

Programs that procure an assistant, deploy it, and subsequently seek evidence of impact produce the outcome described above.

The effective sequence instruments the service, identifies the highest abandonment points, establishes the cause of each, and then determines whether artificial intelligence addresses that specific cause.

In some cases it does. In many cases the resolution is a revised instruction, a removed field, or an integration that eliminates a request for information the government already holds. That determination is substantially more difficult to make after capability has been procured.

Applications that contribute materially

Once abandonment causes are established, candidate applications become specific.

Eligibility pre screening. A meaningful proportion of abandonment occurs after an applicant invests effort and then determines, correctly or incorrectly, that they do not qualify. A plain language indication provided at the outset reduces both applicant burden and downstream review volume.

Plain language translation. Statutory text is precise and largely inaccessible to general readers. Rendering documentation requirements into a description of what an applicant must provide constitutes well suited work, with authoritative language remaining accessible.

Document intake validation. Rejected submissions attributable to incorrect forms, illegible scans, or missing pages generate substantial rework. Validation at the point of upload prevents a rejection cycle measured in weeks.

Status transparency. A significant proportion of contact center volume consists of status inquiries. Improving status legibility reduces call volume more reliably than improving call handling capacity.

Boundaries that should be maintained

Artificial intelligence advises without determining. No adverse determination regarding benefits, eligibility, or legal status should be produced by a model. M-25-21 places systems affecting rights or access to benefits under heightened requirements including pre deployment testing, impact assessment, and ongoing monitoring. Maintaining the model in an advisory position substantially simplifies the compliance position.

Consequential paths retain a human route. Exceptional circumstances occur, and some applicants require direct assistance from staff.

Automated assistance is disclosed. Applicants should be informed when they are interacting with an automated system, and escalation to a person should not require persistence.

Recommended starting point

Instrument the current service, identify the highest abandonment steps, consult the staff who handle the resulting exceptions, and determine what resolves each. The determination should remain open to the conclusion that the resolution is not a model.

The blocking step is consistently more ordinary than teams anticipate. It is rarely the eligibility logic or the identity verification. It is more frequently a document requirement phrased in terms that lead applicants to conclude they lack something they possess, or a field requesting information an applicant has no method of obtaining. Problems of that description are resolved with revised language rather than with a system. For that reason I would allocate the first week of a service engagement to contact center staff rather than to analytics review, as they can generally identify the step before the data demonstrates it.

The measure of a public service is not the sophistication of the interaction. It is whether the individual obtained what they required.