AI-enabled public administration work: use a decision authority map before acting

Define the administrative decision

Write the decision as a specific administrative act: approve a payment exception, change a service rule, prioritize inspections, issue public guidance, or escalate a case. Name the population, programme, jurisdiction, time period, and official who can authorize the act. Separate a recommendation from the decision itself.

The U.S. Office of Management and Budget's M-25-21 memorandum applies to covered U.S. federal agencies and sets governance and risk-management requirements for agency AI use. It also says existing officials retain their legal and policy responsibilities. This is a current primary example of authority remaining with the institution rather than moving to an AI system.

The U.S. Government Accountability Office's AI issue page organizes its accountability framework around governance, data, performance, and monitoring. GAO presents these as practices for federal agencies and other entities, not as a finding that any individual use is compliant or effective.

Sources: [1] [2]

Build the authority map

For each material step, record seven fields:

The map should expose broken links. If an AI draft cites a programme goal but no operative rule, the team still needs the governing source. If a policy permits an action but the analyst lacks delegated authority, the recommendation must be routed to the appropriate official.

• Public purpose: the service, duty, or programme objective the step supports.
• Proposed act: exactly what will change, who will experience it, and when.
• Governing source: current law, policy, delegation, procedure, contract, or approved operating rule.
• Evidence: records and analysis supporting the act, including freshness and known gaps.
• Authority: the role that may recommend, approve, execute, review, and publish.
• Impact check: affected groups, access needs, appeal route, privacy, security, and foreseeable failure.
• Record: decision rationale, version, date, owner, communication, and monitoring trigger.

Worked artifact: a synthetic permit backlog response

A fictional municipal programme team asks an AI assistant to recommend how to reduce a permit-review backlog. The assistant proposes fast-tracking low-value applications and automatically closing files with no response after fourteen days.

The authority map changes the decision. The current operating procedure allows the programme director to reorder work for documented service reasons, but only the appeals unit may close an application. The dataset also records submission date but not whether a notice was delivered in an accessible format.

| Map field | Recorded result | | --- | --- | | Public purpose | Reduce wait time without removing review or appeal rights | | Proposed act | Triage complete low-complexity files into a dedicated queue | | Governing source | Current queue-management procedure and director delegation | | Evidence | Age, completeness, complexity flag, reviewer capacity, missing accessibility confirmation | | Authority | Programme director approves triage; reviewers decide files; appeals unit handles closure | | Impact check | Compare waiting-time distribution and sample accessibility records | | Record | Four-week pilot, weekly exception log, stop trigger for unequal delay growth |

The team rejects automatic closure because the proposed owner lacks authority and delivery evidence is incomplete. It narrows the intervention to a reversible triage pilot, records the rule used, and keeps application decisions with assigned reviewers. The example is synthetic and is not legal advice or a model for any real jurisdiction.

Verify authority, evidence, and recourse

Open the governing source rather than relying on the AI summary. Confirm its status, jurisdiction, effective date, definitions, exceptions, and delegation chain. If several sources conflict, identify which one controls and obtain specialist review when needed.

Reconcile the proposed action with the underlying records. Test missing data separately from negative findings. Compare who receives faster, slower, or different treatment under the option. Identify how a person can correct a record, request accommodation, ask for review, or appeal where those routes exist.

Before execution, assign a disposition of approve, revise, hold, or escalate. A hold should identify the missing authority or evidence and the official able to resolve it. After implementation, compare observed effects with the stated purpose and reopen the map when the rule, data, delegation, population, or operating context changes.

Assess bounded administrative judgment

A controlled simulation can present mixed-version policy documents, a service dataset, an AI recommendation, a delegation table, and a deadline. Observable evidence may include how the participant distinguishes policy from authority, handles missing records, identifies affected groups, routes approval, preserves recourse, and explains the decision.

Report only what the person demonstrated in the scenario and which evidence was available. Do not infer broad public-sector competence, legal compliance, fairness, programme impact, or future job performance from one exercise. Two Prune can organize evidence from a synthetic or approved non-sensitive assessment. It does not issue public decisions, interpret law, certify government systems, or make employment decisions.