Build AI for public services: make the next step easier to find

A public-service employee sits at an accessible information desk beside a neighborhood map and document trays.

Build with AI · Sources reviewed 2026-11-19

A resident should not need to understand an agency’s organization chart to find the right service. AI can help interpret a request, retrieve approved information, and prepare a clear route to the responsible team. The system should make it easier to reach a person when the answer is incomplete or the situation does not fit the standard path.

Start with service navigation or internal preparation. Eligibility, enforcement, benefits, and other consequential decisions require a separate assessment of the applicable law, authority, evidence, and review process. The example here is a fictional U.S. local-government routing workflow.

On this page

First build: a service-request navigator

Create a small public collection with approved descriptions for streetlight reports, missed waste collection, and general permit questions. The assistant proposes the relevant service, links the current official page, and lists information the resident needs to provide. It does not decide an application or promise a response time not stated in policy.

Prerequisites include an approved service catalog, current ownership information, an accessible input and output interface, and a human contact route. Define what can be handled without collecting personal information. Keep the first prototype entirely on fictional requests.

Build the workflow

  1. Accept a short description and an optional broad location needed for routing. Explain which information is needed and avoid inviting unnecessary sensitive details.
  2. Determine whether the request fits the supported catalog. Urgent or unsupported topics should follow the agency’s established contact process.
  3. Retrieve the matching official service page and its current owner. Record the page’s version or check date.
  4. Ask the model to propose a service category, explain why it fits, and identify missing information. Require source links for process statements.
  5. Validate the category against the actual service catalog. A model-invented department becomes a review item.
  6. Present an accessible explanation and the human contact option. For ambiguous requests, show the uncertainty rather than forcing a category.
  7. After the resident or staff confirms, create a request through the approved system. Return the actual receipt or case number. A drafted description is not proof that a service request was submitted.

Example instruction:

Route the fictional request using the supplied public service catalog.
Return a supported category, reason, official link, and missing fields.
Do not decide eligibility, impose requirements, or promise deadlines.
If the catalog does not support an answer, provide the approved
human contact route and say what remains uncertain.

A fictional request

A resident writes: “The lamp beside the park entrance has been dark for three nights.” The catalog distinguishes streetlights maintained by the city from park lighting handled by parks staff. The precise location is absent.

The useful answer asks which entrance or nearby landmark identifies the lamp. It explains that ownership must be checked before routing and links the approved reporting page. It should not invent an asset number or announce that a repair crew has been dispatched.

If the request later becomes a tracked case, the system should preserve the resident’s original description alongside the reviewed summary. Staff can then understand what was changed and avoid losing an important detail during automated rewriting.

A simpler variation rewrites an approved public service page in plain language for editorial review. Preserve the official requirements and contact details, link the original, and have the service owner confirm that simplification has not changed meaning.

Accessibility and accountability are part of delivery

The Department of Justice’s Title II web and mobile accessibility rule addresses content provided by U.S. state and local governments and specifies WCAG 2.1 Level AA as its technical standard, subject to the rule’s scope and exceptions. Check current applicability and compliance dates for the entity. An assistant’s generated content and interaction must be evaluated as part of the service. DOJ rule fact sheet

Design for keyboard use, assistive technology, understandable errors, and alternatives to the chat interface. Evaluate language access with qualified reviewers where offered. These checks should involve people who use the service.

NIST’s AI Risk Management Framework is voluntary guidance for incorporating trustworthiness into AI design, use, and evaluation. It can organize the pilot’s responsibilities and evidence; it does not replace the agency’s legal authority or procurement requirements. NIST AI RMF

Evaluate the route, not just the answer

Test ambiguous locations, requests outside the jurisdiction, obsolete pages, unavailable departments, sensitive personal information, and messages that attempt to override the catalog. Include requests written in plain language and with spelling mistakes.

Measure correct routing, successful contact handoff, source accuracy, accessibility findings, and staff corrections. Check whether people can understand the next action without relying on the assistant alone. Review patterns of misrouting across request types and users rather than celebrating one average accuracy score.

A small municipality may begin with a few public pages and a staffed queue. A county may need department ownership, records handling, and coordinated updates. A larger agency adds procurement, security assessment, retention, service monitoring, and formal review across programs.

Name the catalog owner and the team responsible for unanswered requests. Budget for policy maintenance, accessibility work, human support, and evaluation. Keep the normal service channels operational when the assistant is unavailable.

Continue with AI governance, workflow design, and enterprise implementation.

Share this page