AI Agent Governance for Public Utility Commissions

Public Utility Commissions are beginning to explore AI tools for document review, complaint classification, and rate case support. Public materials describe most current deployments as decision-support tools rather than autonomous agents with independent write access to regulatory systems. That distinction matters. A tool that summarizes a filing for a human reviewer carries different risk than an agent that can query a docket system, cross-reference consumer complaint records, and update a case status without a human approving each step.

As PUCs move toward the latter, governance frameworks built for static software tools or human staff accounts no longer fit. No state or federal PUC has published a governance framework specific to autonomous agents in rate case, tariff, or complaint systems. Commissions moving in this direction are effectively building policy ahead of established precedent, which makes runtime controls, not just upfront approval processes, the practical safeguard.

What AI Agent Governance Means in a Regulatory Context

AI agent governance for a PUC is the set of controls that determine what an agent is allowed to access, what actions it can take, and what record exists of what it did. This differs from general AI policy, which typically addresses model selection, disclosure, or acceptable use. Agent governance is operational: it governs identity provisioning, permission boundaries, and runtime enforcement for a system that can independently invoke tools and touch regulatory data.

Existing federal guidance, including OMB Memorandum M-24-10, directs agencies to maintain AI use case inventories, classify risk, and preserve human oversight for AI that affects rights or safety. Rate case determinations and consumer complaint resolutions plausibly fall into that category. NIST's AI Risk Management Framework and its Generative AI Profile provide risk categories such as data privacy and information integrity that apply directly to agents handling ratepayer and utility filing data. None of this guidance was written for agentic tool-calling architectures, so applying it to PUC AI agents requires extending existing risk-management practice rather than following a ready-made template.

Agent Identity and Permission Architecture

NIST SP 800-53 treats non-human or service accounts as requiring their own identity management, access review, and audit logging, separate from human user accounts. Applied to a PUC environment, this means an AI agent used for complaint triage should not operate under a staff member's credentials or a shared service account. It should have its own identity that can be reviewed, scoped, and revoked independently.

Permission boundaries should be defined per system rather than as a single blanket credential. A docket management system, a billing or complaint database, and a tariff repository each carry different sensitivity levels and different data handling rules. Granting an agent uniform access across all three, because it is convenient during initial deployment, creates a least-privilege violation that NIST's AC-6 control family is specifically designed to prevent. Legacy PUC case management systems were generally not built with granular non-human identity support, which means many commissions will need an identity broker or middleware layer to enforce these boundaries rather than relying on native system controls.

Runtime Policy Enforcement for Tool Calls

Static, upfront permissions are insufficient for agents that make sequential decisions across a session. A policy enforcement layer positioned between the agent and backend systems allows each tool call to be evaluated against defined rules at the moment it is requested, rather than relying solely on access grants set at deployment.

This matters in a PUC context because utility and ratepayer data often carries state-specific confidentiality designations. An agent that has legitimate access to a rate case docket should not be able to combine that data with a separate complaint record in a way that violates a jurisdiction-specific confidentiality rule, even if it technically has read access to both systems individually. Runtime enforcement can apply that boundary at the point of action, where static access control lists cannot.

Audit Logging and Public Records Obligations

Audit logs for PUC AI agents need to capture agent identity, the specific tool or action invoked, the data accessed, and the outcome, at each discrete step rather than only at session start and end. This granularity is necessary to reconstruct a decision pathway if a rate case determination or complaint resolution is later questioned.

State public records laws generally require that records used in government decision-making be retained and available for inspection absent a specific exemption. Whether AI agent logs, prompts, and outputs qualify as public records has not been settled in any specific PUC jurisdiction based on available research, so commissions should determine this in advance rather than after a disclosure request arrives. A related tension is that detailed audit logs may themselves contain sensitive ratepayer or utility data, meaning the logs require their own access controls, not just the systems they describe.

Human Oversight and Accountability Assignment

OMB M-24-10 requires human oversight for AI systems affecting rights or safety, including the ability to intervene or override AI-driven decisions. For a PUC, this translates into defined checkpoints: any agent action that affects a rate case determination, a consumer complaint resolution, or the status of a public filing should require human review before it takes effect, not after.

Accountability for these actions should be assigned to a specific designated official rather than left diffused across IT, legal, and regulatory staff, which is a common failure mode when agent oversight is treated as everyone's responsibility instead of someone's job. Independent evaluation of an agent before deployment, a practice recommended for federal rights-impacting AI, is a reasonable baseline for state PUCs even without a specific state mandate requiring it.