Public Sector AI: Practical Uses With Clear Accountability

Public sector AI can help staff prepare research summaries, draft public explanations, and document administrative processes when the information and tools are approved for that use. Start with a bounded preparation task and a named human reviewer, with decisions and publication handled through the agency’s authorized process.

The practical challenge is carrying the rules and review history from one session to the next. fenbs can keep approved work instructions, tasks, and review decisions available to connected assistants for nonrestricted projects. An agency must approve the service and the information involved before using it for official work.

Which public-sector tasks make a sensible starting point?

Choose something a knowledgeable person can check against a small set of authoritative sources. Rewriting an approved public service page into a draft checklist is easier to review than asking a model to determine eligibility from a resident’s case file. Define the output and the decision it will support before choosing the tool.

A sensible pilot excludes restricted information and does not change anyone’s rights or access to services. That is a starting boundary for the pilot, not a claim that every other use is prohibited. More consequential applications require their own authority, assessment, and safeguards.

The general guide to using AI at work explains how to define a task and review its output. Public agencies also need to establish who has authority over the resulting action.

Who remains accountable for the result?

Name separate owners for the process, the source material, the review, and the final action. One person may hold several roles in a small office, but each responsibility should still be explicit. “The team will check it” does not say who can approve publication or stop a flawed process.

The GAO AI accountability framework↗ (opens in a new tab) organizes accountability practices around governance, data, performance, and monitoring. Use those areas as questions for a pilot: who owns it, what information it uses, how you test it, and who watches for problems after adoption.

For example, a communications specialist might review the readability of a draft while a program owner verifies its meaning. Approval of wording should not silently become approval of a new policy interpretation. Record the scope of each approval and the version it covers.

In fenbs, people decide and sign off; assistants can record what those people chose. Keeping that distinction explicit helps the next assistant recognize an open question rather than treating its own recommendation as authorization.

What information may staff give an assistant?

Use the agency’s approved information-handling rules and review the specific service, account arrangement, and use case. A public document can still be paired with an internal instruction that must stay restricted. Evaluate the complete input, including attachments, copied notes, and any connected sources.

• Begin with published materials or synthetic examples the information owner permits you to use.
• Keep resident identifiers, personnel records, investigation materials, credentials, and other restricted information out of an unapproved service.
• Check whether apparently anonymous details could identify someone when combined.
• Confirm who can access the workspace, what connected tools can read, and how the agency will retain required records.

A vendor’s availability to the public is not agency approval. Likewise, approval for drafting from public materials does not authorize connecting an internal case system. Ask the responsible security, privacy, procurement, and records staff about the particular boundary you want to cross.

Keep the shared instructions precise enough to be useful without including the restricted examples themselves. “Use only published program materials” can guide a pilot without copying an actual resident’s request into the project record.

How do you draft an explanation without changing the policy?

Provide the approved source, identify its jurisdiction and applicable version, and ask for a narrow transformation. Request a plain-language draft with each factual condition traceable to a source passage. Ask the assistant to list unclear points instead of resolving them through guesswork.

Task: turn the supplied approved public guidance into a draft checklist.
Audience: a first-time applicant.
Preserve every condition, exception, deadline, and required document.
Identify the source passage supporting each checklist item.
List conflicts or missing information as questions for the program owner.
Do not invent policy, eligibility rules, or additional requirements.
Return a draft and a review checklist. Do not publish anything.

Review meaning before style. A shorter sentence can be wrong if it removes an exception or changes “may” to “must.” Check the draft in both directions: every claim needs support, and every important source condition needs to survive the rewrite.

Then check the actual resident journey. Can someone identify the next action, the correct contact, and the place to ask for help? Test those steps against the live service rather than assuming a clear-looking paragraph is usable.

What records should the pilot preserve?

Ask records staff to identify what must be captured, where it belongs, and which retention instructions apply before the pilot starts. Useful working information can include the task purpose, source versions, draft, substantive corrections, approval, and the final published item. The official requirements depend on the agency and use.

For federal agencies, NARA’s guidance on AI materials↗ (opens in a new tab) addresses applying the Federal Records Act to inputs, outputs, audit trails, and related materials. It also explains that disposal of AI-related federal records must follow a NARA-approved records schedule. State and local offices need their applicable records guidance rather than assuming the federal schedule applies.

Do not adopt a blanket rule to keep every chat forever or delete every draft immediately. Have the records owner determine the right treatment for the actual use. A task board’s activity history does not by itself establish that an agency has met its records obligations.

An outside AI brain can help keep the work understandable between sessions. It should point staff toward the approved record location rather than becoming an accidental substitute for the agency’s recordkeeping process.

How should accessibility and errors be reviewed?

Review the delivered format as well as the words. A correct draft in an inaccessible PDF or a form with unclear labels can still block access to a service. Include people who understand the audience and test the real publication format with appropriate accessibility methods.

The Justice Department’s state and local web accessibility fact sheet↗ (opens in a new tab) explains the rule covering web content and mobile apps provided by state and local governments. Consult the agency’s accessibility owner about applicable requirements and exceptions instead of treating an AI readability check as proof of compliance.

Use a review set that includes unusual but valid situations, contradictory materials, missing information, and requests outside the task scope. Check whether the assistant makes up an answer, omits an exception, or points to an unusable contact route. Log the failure and the correction without copying restricted details into a broader workspace.

Make correction ownership explicit after publication too. If a source changes, someone must identify affected material and review it. Saving an approved prompt does not automatically update the pages that were produced with it.

How to do this with fenbs

1. Create a free account for an approved, nonrestricted exercise, or use synthetic material to evaluate the workflow before any official use.
2. Add a project with the narrow purpose, permitted inputs, responsible reviewer, and official record location. Create tasks for drafting and review.
3. Record the rules a person has approved in Decisions and rules, including the source boundary and who may authorize publication.
4. Connect a supported assistant and ask it to read the context before preparing a draft. Have the responsible people review the result in the approved process.
5. Save progress with completed checks, open questions, and the next step. Preserve required official records in their designated location.

fenbs is free for one person, with no credit card. It is not the right fit for restricted casework, an official records repository, or an automated eligibility decision system. Do not use it for agency information unless the agency has approved that use.

When is the workflow ready to expand?

Compare the pilot with the existing process using measures people can verify. Track substantive corrections, reviewer effort, missing conditions, and whether the final explanation helps someone complete the intended step. Count unsuccessful or abandoned drafts too, so the comparison reflects the whole job.

Expand only after the owner approves the next scope and the supporting information, review, records, and correction processes are ready. Record what the pilot demonstrated and what it did not test. An accurate public-information checklist does not establish readiness for resident-specific advice or decisions.

For nearby topics, see AI tools available to the public and AI in public health work; more professional workflows are in the role guides.