Utilities & Critical Infrastructure
AI Agent Governance for Public Safety Power Shutoffs
Governing AI agents in PSPS workflows requires scoped, task-specific permissions rather than broad system access, mandatory human approval before any shutoff-influencing action, distinct and revocable agent identity separate from human operators, and audit logs that capture input data, model version, and approval status for every decision the agent contributes to.
Runtime Governance
Least-Privilege Agent Identity and Tool-Call Scoping
These four controls define how AI agents can safely participate in PSPS decision workflows without acquiring authority beyond their intended scope.
Scoped Tool Access
Agents restricted to pre-approved functions rather than open-ended queries or commands.
Human-in-the-Loop Gates
Mandatory approval checkpoints before any recommendation with public safety consequence.
Distinct Agent Identity
Machine identities separate from human credentials, independently revocable.
Replayable Audit Trail
Decision-path logging sufficient for post-event regulatory review.
Where AI Agents Are Entering PSPS Decision Workflows
Utilities are integrating AI agents into stages of the Public Safety Power Shutoff process that were previously manual: correlating weather forecasts with grid telemetry, scoring wildfire risk against asset and vegetation data, and drafting shutoff recommendations for operator review. These agents typically pull from multiple systems of record, including real-time telemetry feeds, weather models, and asset risk databases, in order to produce a single risk assessment or recommendation. This is a multi-source pipeline, and each data source an agent touches represents a distinct access boundary that needs to be defined rather than inherited from broad system credentials. The governance question is not whether AI can assist with PSPS analysis, but what boundaries constrain what the agent is permitted to read, compute, and recommend, and where a human is required to intervene before any action with real-world consequence occurs.
Human-in-the-Loop Checkpoints: Mandatory Versus Advisory
Not all human review is equivalent. A checkpoint that a human operator can skip, override without documentation, or that occurs after an action has already taken effect provides a different level of control than one that blocks execution until approval is recorded. Given the public safety and liability consequences of a PSPS decision, governance frameworks should explicitly state which stages of the workflow allow agent autonomy and which require mandatory, logged human approval before the recommendation proceeds. Ambiguous framing, where a checkpoint is described as human oversight but functions as an optional review, misrepresents the actual level of control in place. Utilities should be able to state plainly, for each stage of the PSPS pipeline, whether the agent's output is advisory input to a human decision or a gated action that cannot proceed without recorded sign-off.
Audit Trails and Regulatory Alignment
High-consequence automated decisions generally require an audit record that captures the input data available at the time of the decision, the model or ag