AI Agent Governance for Public Housing Authorities

Why This Requires Deliberate Governance, Not Default IT Practice

Public housing authorities are starting to deploy AI agents for tenant communications, eligibility screening, maintenance triage, and internal case management. These use cases touch resident income data, eligibility determinations, and case records that are already subject to HUD data security requirements. No HUD-specific regulation currently addresses AI agents directly. This means PHAs must interpret and apply existing frameworks, including HUD Handbook 2400.25, EIV security protocols, and general federal standards like NIST SP 800-53, to a category of system those rules were not originally written for. Treating agent deployment as a standard software rollout, without adapting access control and audit practices to how agents actually operate, creates a compliance gap that will surface during an audit rather than during design review.

The Legacy System Problem

Most PHA case management systems, including PIC/EIV-connected tools and older tenant portals, were not designed with API-level granular access control. They typically support user login and role-based access at a coarse level, sufficient for a caseworker logging in to view a file, but not for scoping what a specific automated action can read, write, or invoke. When an AI agent needs to summarize a tenant communication, check eligibility status, or update a maintenance ticket, the underlying system usually cannot natively enforce which of those actions the agent is permitted to take. This is a structural constraint, not a policy oversight, and it means governance controls generally need to sit above or alongside these systems rather than rely on native enforcement.

Agent Identity Must Be Distinct From User Identity

A recurring gap in older role-based access control models used by PHAs is the absence of a clear distinction between an end user's authority and an AI agent's authority acting on that user's behalf. An agent should not simply inherit a caseworker's full permission set. Instead, agent identity should be provisioned separately, with scoped and time-limited permissions tied to specific task types. An agent handling maintenance triage should not carry the same access as one processing income verification data under EIV rules. This distinction matters directly for compliance, since EIV already mandates role-based access control and user-level accountability for anyone, or anything, touching income and eligibility data.

Runtime Enforcement as the Practical Layer

Because legacy PHA systems cannot enforce task-based permission scoping natively, runtime policy enforcement between the agent and those systems becomes the practical mechanism for closing the gap. This layer can apply least-privilege rules at the point of each tool call, distinguishing read-only actions such as summarizing communications from write actions such as modifying eligibility or case data. Write actions carry materially higher compliance risk and should be subject to stricter scoping, approval workflows, or human review, particularly given Fair Housing Act and Section 504 obligations around nondiscriminatory outcomes in tenant selection and eligibility determinations. Runtime governance platforms are one way to implement this enforcement layer without requiring PHAs to build bespoke controls into each connected system, which is generally not feasible for resource-constrained IT teams.

Audit Logging That Satisfies Oversight, Not Just Uptime Monitoring

Standard user-login audit trails, common in legacy PHA systems, are not sufficient to demonstrate accountable AI agent behavior during a HUD or state oversight review. Audit logging for agents needs to capture the inputs that drove a decision, the specific tool calls invoked, and the data accessed, tied to the policy and permission set in force at the time of the action. This is a materially different logging requirement than session-level activity tracking. PHAs should treat this as a design requirement from the outset of agent deployment, since retrofitting action-level logging after an agent is already in production is significantly harder than building it in from the start.

Evaluating Governance Approaches

When assessing how to govern AI agents, PHA leadership should evaluate whether a proposed approach can enforce task-scoped, time-limited permissions against systems that lack native API controls, whether it generates action-level audit logs sufficient for HUD or state review, and whether it separates agent identity from end-user credentials in a way that supports revocation per task type. These are governance mechanics, not abstract AI ethics questions, and they determine whether an agent deployment can withstand an audit. Runtime governance platforms, including Trussed AI's runtime policy enforcement and audit logging capabilities, are one practical way to apply these controls across legacy systems without requiring custom engineering for each integration, but the underlying requirements exist independent of any specific vendor.