Government administrators are experiencing rapid policy changes at the federal, state, and local levels. These changes force them to update rules, procedures, training, oversight, and communications to comply with executive orders, statutes, and regulations while managing day‑to‑day operations under tight timelines and constrained staffing. They may also face multiple layers of legal and budget review before new documents can be finalized. In this environment, agencies need practical tools to accelerate policy analysis and document updates without sacrificing accuracy, transparency, or care.

For example, state Medicaid agencies often must adjust procedures to implement new federal directives around work requirements and changes to the Affordable Care Act—rewriting eligibility manuals, notices, standard operating procedures, and training, and communicating clearly with beneficiaries and partners in a matter of months.

One solution to speed the update of existing agency documents to comply with new directives is using a private, generative artificial intelligence (GenAI) assistant. The GenAI can ingest the new directives alongside the agency’s current manuals and forms, pinpoint the sections affected, propose draft redlines and plain‑language notices with citations, and produce an “old vs. new” cheat sheet for field staff.

A private, agency‑document GenAI assistant reads only documents the agency provides—it does not search the internet. Agency teams refine drafts it creates—the GenAI assistant simply gives them a head start, keeping every change traceable while preparing documents for leadership decisions. The result is a coordinated, auditable rollout that provides higher-quality materials to meet an administrator’s tight deadlines.

Pulling information from the internet can surface outdated, conflicting, or irrelevant guidance that does not reflect local or current policies. Even when an AI assistant is restricted to internal documents, large language models can still “guess” if not governed properly. Two solutions address this challenge: limiting the AI assistant to only approved sources and requiring citations for substantive claims. Together, these reduce AI hallucinations and keep answers grounded in agency source documents.

Using a private, agency‑document GenAI assistant to analyze policy changes provides tangible value. It can: 

• Compare old and new statutes, regulations, and administrative rules
• Flag conflicts, dependencies, and preemption issues that affect local procedures
• Generate checklists to update policies, forms, training, timelines, and reporting
• Produce plain‑language briefs for leaders and staff with citations to the source text

A simple design makes this process workable. Staff can limit the GenAI assistant’s knowledge base to ensure its answers will only come from agency documents. If it can’t find a supporting passage, it will say so. Agencies can also require the tool to list citations, so every substantive answer points to the exact page or section of the document where it found the answer.

To begin, agencies need:

• Subject matter experts who know the policies and authoritative sources
• Analysts or records managers who organize documents and metadata
• A light‑touch technical administrator to manage the documents and configure the repository, access controls, and vendor tools—requiring minimal coding

Having users construct clear, well‑structured requests helps. These can include defining the task, naming the sources to use, specifying the desired output, and setting guardrails such as requiring citations and applying abstain rules when sources are insufficient or conflicting.

• Gather all relevant public or publishable documents—plans, policies, ordinances, board packets, and finalized guidance.
• Build a searchable, versioned repository and disable web access for the assistant, allowing only approved sources.
• Pilot an assistant that requires citations, includes explicit abstain rules, and logs sources, decisions, abstentions, and changes. Note that it’s important to keep humans in the loop so the AI drafts documents and staff review and approve them.
• Measure what matters—accuracy, time saved, staff effort, abstain rate, and stakeholder satisfaction—and iterate and scale from there.

The first step of identifying the right tools for an agency-document Gen AI assistant is critical. Constraining an LLM to focus on a limited set of documents is possible with all LLMs, but may be easier with some that are specifically designed to do so, like Google’s NotebookLM. NORC has both the technical and subject matter expertise to assist state and local agencies with identifying tools, establishing the necessary guardrails for success, and using GenAI to conduct policy analysis.

• Administrators need to update policies and procedures on tight timelines with limited staff and funding.
• Private, agency‑document GenAI assistants can help administrators who have limited time, staff, and funding by digesting regulatory documents and drafting new policies and communications quickly while remaining source‑grounded.
• Success depends on strong subject matter expertise, good document organization, and clear task definitions, not heavy coding.
• Start with policy impact analysis on public or publishable documents, require citations, and set explicit abstain rules with escalation to staff.