Government AI Administration Plan

Employees remain responsible for decisions, communications, and work products created with AI assistance.

AI use must comply with public records, privacy, civil rights, accessibility, procurement, records retention, cybersecurity, and employment laws.

Oversight increases when a system affects rights, benefits, safety, legal status, employment, enforcement, or access to public services.

Sensitive information may be used only in approved systems with appropriate contractual and technical protections.

The municipality documents approved uses and provides public notice when disclosure is legally required or important to public trust.

Every system must have an exit plan, a human fallback, and access to authoritative records outside the AI tool.

The department identifies the problem, proposed tool, users, affected residents or employees, expected benefit, alternatives considered, data involved, integrations, cost, and department owner. Upgrades that add AI to an existing system also require review.

The AI Administrator confirms whether the proposal is actually AI, checks for duplicate tools, assigns a preliminary risk level, and identifies required reviewers.

The Review Team evaluates legal authority, public records, retention, privacy, cybersecurity, accessibility, equity, data quality, accuracy, human oversight, vendor terms, data location, model training practices, subcontractors, incident response, and exit options.

The decision is documented as approved, approved with conditions, returned for revision, or rejected. Approval states the permitted purpose, users, data, controls, testing period, disclosures, owner, and review date.

The department tests the system with representative users and realistic scenarios. Testing includes error rates, harmful bias, prompt injection or misuse, accessibility, records capture, security, and fallback procedures.

Production access is limited to trained users. Required notices, support materials, logging, escalation procedures, and manual alternatives are in place before launch.

The owner tracks performance, complaints, incidents, overrides, changes in vendor terms or models, and whether the system continues to deliver the approved benefit.

Approval is reviewed at least annually and whenever the use, data, integration, model, vendor, or law materially changes.

• Use only municipality-approved accounts, tools, integrations, and browser extensions for government work.
• Block or warn against entry of Social Security numbers, payment-card data, credentials, protected health information, criminal justice information, confidential legal material, and other restricted data unless the system is expressly approved for that information.
• Apply least-privilege access, multifactor authentication, logging, retention settings, and periodic access reviews.
• Require employees to verify facts, sources, calculations, citations, tone, accessibility, and legal sufficiency before relying on or publishing AI output.
• Do not allow AI to make an unsupervised final decision that materially affects a person’s rights, safety, eligibility, employment, enforcement status, or access to services.
• Maintain records needed to explain the purpose, inputs, responsible officials, material outputs, decisions, testing, incidents, and changes associated with the approved use.
• Provide a clear way for residents and employees to reach a person, correct inaccurate information, submit feedback, and appeal consequential decisions.
• Stop or restrict the system when an incident, material error, contract change, model update, or legal change creates unacceptable risk.

Maintain an internal inventory of every approved AI use and consider publishing a public-facing summary for systems that interact with residents or materially affect services.

At minimum, the summary should explain the system’s purpose, responsible department, data categories, human oversight, review date, and how to ask questions or challenge an outcome.

Public-facing deployments should include plain-language education, accessible alternatives, and a feedback channel.

For controversial or high-impact uses, consider a resident advisory group, public meeting, or time-limited pilot before full deployment.

1. Stop or isolate the affected use when continued operation could cause harm.
2. Preserve relevant prompts, outputs, logs, records, configurations, and vendor communications.
3. Notify the AI Administrator, security, legal, privacy or records officials, leadership, and other required parties.
4. Assess the scope, affected people, data exposure, decision impact, and required notices.
5. Correct records or decisions, provide a human review or appeal, and communicate with affected people when appropriate.
6. Document root cause, corrective action, lessons learned, and whether the system may resume.

Framework note: This plan reflects the NIST AI Risk Management Framework’s Govern, Map, Measure, and Manage functions, federal public-sector practices for AI inventories and risk management, and municipal guidance emphasizing human review, privacy, transparency, approved procurement, and documented risk levels. It should be reviewed by the municipality’s legal counsel and records, privacy, cybersecurity, procurement, accessibility, and human resources officials before adoption.