Constituent data stewardship in public-sector AI solutions

An AI adoption plan must support an organization’s specific needs, and first steps require research—determining what options support resilient foundations, intelligent systems with guardrails, and secure data.

For those working in the public sector, AI deployment needs to consider security integrations and preservation of data ownership. State and local government workers certainly need faster information access, but handling sensitive constituent data (personally identifiable information, or PII) requires data and records protection. Digital administrators need to develop documentation and policies that answer complex questions about information processing and data retention.

How does government provide streamlined public services while keeping constituent data private and secure?
What actions can prevent constituent data from being used to train AI models?
How can government block third-party transmission of constituents’ web searches or prompts?

The first question is perhaps the most confusing: Where do I start?

Step 1: Place AI governance within data stewardship

Framing AI governance in terms of compliance and data stewardship will help make a case not only for worker efficiency, but also for resilient systems that prioritize constituent privacy.

Frame AI within methodical data stewardship.
Identify risks to avoid harm.
Define security priorities.
Set applicable standards and quality checks for AI governance.
Underscore that deployment design determines whether prompts, files, searches, and logs stay under government control.
Support proposed frameworks / AI standards with government, academic, or nonprofit primary sources, such as the EPI Center guide, which recommends that AI be integrated through mission and program structure, with support from all stakeholders.

Start with RAG

Retrieval-augmented generation (RAG) is an AI framework that uses the combined methods of traditional information retrieval (such as databases and search) and large language models (LLMs). RAG is a way for AI to retrieve approved documents before answering, which grounds outputs in sources that government supports or owns.

RAG’s method of providing context poses a useful starting point—it can improve responses using approved documents without requiring model retraining on sensitive records.
Government ownership over constituent data is integral to security; RAG only works safely if the government controls retrievals: the document store, access permissions, and logging.
Refer to NIST’s framing that AI risk management should be built into the design, development, use, and evaluation of AI systems—not added on later.

Establish definitive guardrails for agents

Agentic AI systems are more autonomous, setting intermediate objectives and implementing ongoing processes. Because they act by structuring “how judgment is exercised rather than merely informing choices,” agents can be useful but also introduce new governance and security concerns.

While agents can streamline multi-step tasks, the same capability exposes more risk, especially if they can access too many tools or move data unchecked across systems.
Request that IT constrain agents to approved tools; build in human-in-the-loop checks and balances at the start of a decision chain.
Humans must validate and review outputs pulling from optimization logic, constraints, and upstream objectives. Consider adding observability tools such as Langfuse or Arize AI for response quality monitoring and anomaly detection.
- Is human approval required for sensitive actions?
- Is every step auditable?
- Has documentation been created for repeatable processes?

Take extra measures to secure web search

Because web search is a feature that may send constituent queries outside a government’s digital environment, it must be evaluated for retention, logging, and third-party transmission risk. Web search is therefore the feature that, when left unchecked, is most likely to transmit constituent queries to third-party services or data-processing centers.

When building and maintaining best security practices, ask whether these searches are retained, logged, or used for training by an outside provider, and whether the government can disable or proxy the feature.
Advocate for a framework that minimizes identified misuse / data-retention risks.
To meet the needs of both government staff and the public, AI governance must center transparency, accountability, privacy, and cybersecurity from the beginning of a modernization initiative.

Step 2: Lead with an integrated security approach

Minimizing data breaches or quiet transmissions to third parties involves reducing the vectors whereby prompts, uploads, search terms, and outputs can leave the controlled digital environment.

Propose a tailored security posture lens to colleagues using the National Institute of Standards and Technology (NIST) AI risk-management framework: organizations should better manage risks to individuals, organizations, and society across the design, development, use, and evaluation of AI systems.
Emphasize that a system can be “useful” and still be unacceptable if it weakens privacy, retention, auditability, or data control.
Seek FedRAMP-compliant LLM hosting through AWS Bedrock or Azure OpenAI, rather than going directly to vendor APIs. If constituent data can be kept within that cloud boundary, spillage becomes much less likely.

Document answers to data-retention questions before AI deployment

Lead discussions on AI governance with curiosity and overt communication. The same questions may require cross-functional collaboration and iteration stages—following up over time, testing out options, and approaching challenges from different angles.

Where do prompts, uploads, and search queries go after entry?
Can the system guarantee that constituent data isn’t used for training?
How will the government control retention, deletion, and logging?
Can web search be disabled or restricted to approved sources?
How will retrieval and agent actions be audited?
Will the deployment run in a government-controlled or dedicated environment with government-owned keys and identity controls?

Step 3: Refine the adoption pitch for IT decision makers

Become an informed participant and partner with organizational technical decision makers. Contextualize AI governance by grounding recommendations and research findings in already approved government guidelines and mission objectives. Explain that identifying and budgeting for AI risk helps to avoid liability and better assess vendor claims.

Share established AI risk-management frameworks from official government sites. Review the AI governance efforts of different states.
Explain that open-source and localized architectures can support stronger control because the government can inspect, host, and govern more of the stack itself. Present the stewardship principle early and clearly: constituent data should remain under local governance, with state- or local-defined retention and access rules.
Drive guidance documentation in terms of expanding access to knowledge while keeping constituent data inside a trust boundary owned by state or local government.
After reviewing results from the pilot, revise future approaches and expansions with insights from privacy, legal, security, procurement, records management, and accessibility.
Close by reinforcing that more features don’t automatically mean better outcomes for all involved. The best public-sector AI deployments offer a helpful user experience while preserving control, accountability, and public trust.