Designing public-sector agents that prepare evidence without deciding eligibility

Begin with the workflow vocabulary

Citizen information, Application completeness checks, Administrative case preparation each have distinct inputs and review points. A shared vocabulary should explain what the core entities—Service, Application, Required document, Procedure version—mean and how they relate. A model cannot repair conflicting business definitions just by receiving a longer prompt.

Preserve context that changes the answer

Preserve procedure versions, separate public and restricted sources, and represent missing information explicitly. Record the source authority and the conditions under which information applies. An answer can be fluent and still be wrong because it used the wrong entity relationship, document version, unit, or time period.

Make evidence a product feature

A casework team checks whether an application has the required documents and prepares a review packet for an accountable official. The user experience should expose the evidence behind a proposed output and make uncertainty visible. Provide source links, relevant dates, and a clear distinction between observed records and model-produced interpretation.

Give agents a bounded responsibility

Retrieve approved service guidance and prepare cited answers. Define this responsibility with explicit completion criteria and a small tool set. Keep retrieval, draft preparation, and record-changing actions separate in the permission model. An agent’s ability to suggest a next step does not imply authority to execute it.

Test the difficult cases first

Include conflicting identifiers, missing evidence, obsolete documents, and restricted records in evaluation. Proposed measures include checklist accuracy, source-version correctness, access-boundary testing, and review packet completeness. Also test whether reviewers can understand and correct an output without reconstructing the whole task themselves.

Keep the ownership model visible

Eligibility, enforcement, and other consequential decisions remain with accountable officials. Data stewards resolve source meaning and quality, platform teams operate access and reliability, and workflow owners accept the business behavior. Treat these responsibilities as part of the architecture rather than a handover note.