AI Agents on Power Grids and the EU AI Act

RestoreBench: The First Public Measurement of AI Agents on Power Grids

On August 31, 2026, six researchers deposited a benchmark on arXiv called RestoreBench. The document carries the reference arXiv:2609.00384v1^[1].

The benchmark evaluates Large Language Model-based agents on a specific task: restoring power flow convergence in electrical grids. It covers two grids and 46 cases per grid, for a total of 92 convergence failure scenarios.

The measurement compares three architectures: chatbot, single agent and multi-agent. Each receives the same simulation environment, the same observation and action spaces, and the same evaluation metrics.

This is the first relevant point. The question about AI agents in critical assets remained qualitative. Now it becomes quantitative and reproducible.

From "Can We Deploy?" to "Who Decides?"

Public coverage of these systems tends to focus on capability. The real transformation concerns accountability.

A reproducible benchmark makes an AI agent's performance falsifiable. It therefore makes regulable what previously remained engineering opinion. Operators of electrical grids now have a common measure.

The shift is clear. Before August 31, 2026, the discussion concentrated on the technical feasibility of deploying agents on critical infrastructure. Now the focus moves to the accountability framework that names the person responsible for the agent's decision when it fails.

A scientific artifact that measures performance opens a legal question. Who is responsible for an incorrect corrective action on a live grid? The answer belongs to governance, before it belongs to technology.

The Regulatory Context: The EU AI Act and Critical Infrastructure

The EU AI Act has been in force since August 1, 2024, with staggered implementation across multiple deadlines. The jurisdiction is the European Union.

Annex III classifies AI systems for managing and operating critical infrastructure, including electrical grids, as high-risk. Obligations for high-risk systems apply by August 2, 2026.

This positions RestoreBench at a precise point in the timeline. The benchmark arrives just weeks before the deadline that operationalizes high-risk obligations. An operator evaluating AI agents for power flow now operates within the Annex III perimeter.

The European Commission has discussed deferrals and simplifications through the Digital Omnibus package. The calendar remains anchored to the text in force.

The Governance Signal

The governance signal: a measurable task becomes an attributable task. When performance is quantified, the absence of a named responsible party becomes visible.

A compliance posture calibrated on "decision support system" logic proves today overcalibrated for a context where the agent executes corrective actions on live assets. The difference matters in audit.

The three tested architectures carry distinct risk profiles. A chatbot suggests. A single agent plans and executes. A multi-agent system distributes decision-making among components. Each architectural choice shifts where human responsibility enters the cycle.

This is the heart of this desk's position: accountability without a name remains compliance theater. A framework that avoids naming a specific role produces documentation, not real governance.

Named Accountability: The Question the General Counsel Must Resolve

The measurement places the burden on the General Counsel and Chief Compliance Officer. The question becomes operational.

Which named role within the organization is responsible for the corrective action executed by an AI agent on power flow, by name, in writing, before deployment? The answer belongs to the accountability register, before it belongs to the model.

Organizations without a designated responsible party face growing exposure. Audit remains required. The audit perimeter changes.

A public benchmark also provides a reference point against which auditors can compare declared performance. Organizations that now build audit trails, risk classification and named accountability gain an 18-24 month advantage when enforcement truly begins.

The Risk Framework to Update

The Chief Risk Officer inherits a precise task: updating the operational risk framework in light of now-measurable performance.

A risk model that treated AI as a support tool underestimates the agent that executes actions. Annex III's high-risk classification imposes specific controls: data management, traceability, human oversight, robustness.

RestoreBench offers useful categories. The 92 instances of convergence failure describe real stress scenarios on two grids. The risk framework gains precision when it incorporates these scenarios as internal test cases.

Regulatory fragmentation complicates the picture. Jurisdictions diverge in pace and content. An operator active on multiple markets builds a framework that absorbs the strictest rule, rather than chasing every local variant.

Three Decisions for the Board

The board of directors faces three concrete decisions, all prior to deployment.

First decision: designate in writing the role responsible for each action executed by an AI agent on critical infrastructure. The name precedes the code. This choice responds to Annex III logic.

Second decision: select the architecture, chatbot, single agent or multi-agent, based on acceptable risk profile, and document the rationale. RestoreBench provides the technical measure for this evaluation.

Third decision: define the disclosure that the Board Audit & Risk Committee brings to stakeholder attention. What residual risk remains after agent adoption, and which human control oversees it?

Regulatory Horizon

Current state: The EU AI Act is in force, with high-risk obligations applicable by August 2, 2026 in the European Union. Annex III includes critical infrastructure.

RestoreBench remains a scientific artifact, deposited on arXiv on August 31, 2026 and also available via alphaXiv^[2]. It carries probative force rather than legal force. It nonetheless becomes a technical reference that auditors and regulators can cite.

The question about whether AI agent performance on electrical grids is measurable has received an answer. A second question has opened: on what accountability framework does the operator name the person responsible for the decision. Organizations that respond now, in writing, arrive prepared for the deadline.