The Story of GridSafe
How a grid operator discovered their AI could literally switch off the power
Fictional scenario — based on realistic situations
The Trigger
How it started
GridSafe had used AI for grid management for years. The system predicted failures, optimized load balancing, and sometimes made autonomous decisions to prevent grid overload. It worked excellently — until the AI Act came.
Critical infrastructure is explicitly named as high-risk in the AI Act. AI systems that can disrupt essential services — energy, water, transport — fall under the strictest requirements. And GridPredict AI actually had the capability to shut down parts of the grid.
"If our AI makes a mistake, hundreds of thousands of households sit in the dark." The technical director only realized the implications when he read Annex III.
The Questions
What did they need to find out?
Why is energy AI critical infrastructure under the AI Act?
The team went through the legal text. Annex III explicitly mentions AI systems intended for use as a "safety component" in the management of critical infrastructure, including electricity supply.
💡 The insight
The rationale is clear: an error in energy AI can be society-disrupting. Blackouts affect hospitals, traffic lights, refrigeration. The AI Act recognizes that AI in this context is as critical as the physical infrastructure itself.
🌍 Why this matters
Europe has experienced multiple large-scale blackouts. The increasing complexity of the grid — with decentralized generation, EV charging, and dynamic pricing — makes AI essential but also risky. Regulators like ACM and RDI follow this closely.
What controls do we need to build in?
GridPredict AI sometimes made autonomous decisions: it could shut down transformers, redistribute load, or activate emergency protocols. The question was: which of these actions should require human approval?
💡 The insight
The team developed an impact matrix. Reversible actions with low impact could be automatic. Actions affecting more than 1000 households required human confirmation. Emergency shutdowns needed two-person approval — except in immediate danger.
🌍 Why this matters
The AI Act asks for proportional human oversight. This doesn't mean AI can never act autonomously, but that controls are proportionate to risk. For critical infrastructure, that bar is high.
How do we guarantee human oversight on real-time decisions?
The dilemma was timing. Some grid problems require response within milliseconds. No human can decide that fast. But the AI Act asks for human oversight. How do you solve that?
💡 The insight
The solution was a combination of pre-approval and post-hoc review. Human operators set the parameters within which the AI could operate. Every autonomous decision was logged for daily review. Deviations trigger alerts.
🌍 Why this matters
This is a fundamental tension in AI governance: speed versus control. The AI Act recognizes this by focusing on "meaningful" oversight — not necessarily real-time, but effective. It's about the system, not every individual decision.
What if the AI causes a cascade effect?
The nightmare scenario: the AI makes a decision that, through unforeseen interactions, causes a cascade of failures. A local shutdown becoming a regional blackout. How do you mitigate that?
💡 The insight
The team implemented "blast radius" limits. Every AI decision had a maximum impact it could cause. On top came an independent monitoring system that evaluated AI decisions in real-time and could override if patterns indicated cascades.
🌍 Why this matters
Cascade effects are a recognized risk in complex systems. The AI Act requires robust risk management, including scenario analysis of worst-case situations. For energy infrastructure, this means collaboration with regulators and grid operators across Europe.
The Journey
Step by step to compliance
The scope analysis
An external audit identified three AI systems falling under Annex III. GridPredict AI was the most critical.
Developing impact matrix
The team categorized all possible AI actions by impact and reversibility. This formed the basis for the governance framework.
Human oversight redesign
The control system was revised. Which decisions could the AI make, which required human approval?
Blast radius implementation
Technical limits were built in preventing any single AI decision from having too much impact.
Monitoring system
An independent system was implemented that monitors AI decisions in real-time and detects anomalies.
Regulator engagement
GridSafe proactively shared their approach with ACM and the Ministry of Economic Affairs. Transparency built trust.
The Obstacles
What went wrong?
✗ Challenge
AI sometimes needed to react faster than humans can decide
✓ Solution
Pre-approval parameters with post-hoc review and anomaly detection
✗ Challenge
Risk of cascade effects from AI decisions
✓ Solution
Blast radius limits and independent real-time monitoring
✗ Challenge
No clear standards for AI in critical infrastructure
✓ Solution
Proactive collaboration with regulators to develop best practices
The AI Act forced us to think about scenarios we'd rather ignore. Our system is now safer — not despite regulation, but because of it.
The Lessons
What can we learn from this?
Energy AI is critical infrastructure
AI that can disrupt essential services falls under the strictest AI Act requirements.
Speed and control can coexist
Pre-approval parameters with post-hoc review provides meaningful oversight without real-time bottlenecks.
Limit the blast radius
Technical limits on AI impact prevent single decisions from having catastrophic consequences.
Collaborate with regulators
Proactive engagement builds trust and helps shape standards.
Does your AI manage critical infrastructure?
Discover what the AI Act means for energy, water, transport and other essential services.
Ga verder met leren
Ontdek gerelateerde content