Eligibility AI Is Classified as Rights-Impacting, Not Routine IT
Federal guidance issued over the past year treats AI systems that determine or influence eligibility for public housing and benefits programs as rights-impacting AI, a classification that triggers the highest tier of required safeguards rather than standard enterprise IT oversight. OMB Memorandum M-24-10, issued March 28, 2024, directs agencies to conduct impact assessments, provide independent evaluation before and during deployment, and maintain a public inventory of AI use cases under an agency Chief AI Officer. Executive Order 14110 similarly directed agencies to address AI risks in programs that affect access to critical government services and benefits. For governance leaders overseeing housing authorities or benefits agencies, this means eligibility AI cannot be governed through model accuracy testing or a one-time compliance review alone. It requires continuous, verifiable controls over how the system behaves once it is operating on live cases, since a decision that appears sound in testing can still act outside intended bounds in production.
Runtime Controls: Agent Identity, Least Privilege, and Tool-Call Governance
M-24-10 requires human oversight, the ability to override AI outputs, and ongoing monitoring for disparate impact, but it does not specify how agencies should technically enforce these requirements at the system level. The closest authoritative reference is NIST Special Publication 800-207, Zero Trust Architecture, which defines least-privilege access and continuous verification as baseline principles for identity and access management on systems supporting automated decision-making. Applied to eligibility systems, this means an AI agent that queries applicant records or an eligibility rules engine should operate under its own distinct, verifiable identity rather than inheriting a caseworker's credentials. Access should be scoped at runtime to only the data fields and tools required for a specific determination, and requests outside that scope should be blocked at the point of the call, not merely flagged in a later review. Runtime policy enforcement of this kind complements pre-deployment testing rather than replacing it, since testing cannot account for every combination of case data and agent behavior the system will encounter in production.
Auditability for Due Process and Appeals
M-24-10 requires agencies to notify individuals when AI is used in a rights-impacting decision and to provide a means to appeal adverse outcomes. Meeting this obligation depends on audit trails that can reconstruct, for a specific applicant, which data fields, eligibility rules, and model outputs contributed to the determination. These records need to be interpretable by caseworkers and administrative law judges, not only by engineering staff, since appeals are typically adjudicated by non-technical personnel working from case files rather than system logs. The same monitoring requirement extends to disparate impact: agencies must retain records that allow AI-influenced outcomes to be correlated with case data, and where legally permissible, demographic data, over time. This is an ongoing operational requirement rather than a pre-launch checklist item, and it depends on logging infrastructure that captures each tool call and data access at the point it occurs, not just the final output of the decision.
Operational Implications and Open Gaps in Current Standards
Two gaps are worth acknowledging directly. First, no federal statute or standard identified in recent guidance names AI agent identity or tool-call governance as discrete regulatory requirements for benefits systems; these controls are inferred from adjacent zero-trust and AI risk-management frameworks, including NIST AI RMF 1.0 and the July 2024 Generative AI Profile. Agencies should not wait for a named standard before implementing them, given that the underlying due process and disparate-impact obligations already apply today. Second, human oversight requirements are only meaningful if override capability is operationally exercised, not documented as a policy statement. This requires coordination between IT security teams and program policy staff, since a change to eligibility rules typically requires a corresponding update to an agent's permission scope and to the audit fields captured for that decision type. Runtime governance platforms that provide agent identity, least-privilege permissioning, tool approval workflows, and audit logging address this operational layer, sitting alongside the impact assessments and human-in-the-loop reviews that M-24-10 already requires. Trussed AI provides runtime governance and security controls in this category for enterprise and government AI agent deployments.