Building confidence through human-AI collaboration, explainable decisions, and operator-first design in safety-critical infrastructure. This page states our EU AI Act posture, how humans stay in the loop, what gets logged, and where your data lives.
AI systems used as safety components in the management and operation of critical infrastructure — including electricity — fall under Annex III, point 2 of the EU AI Act, and the high-risk obligations have been enforceable since 2 August 2026. We treat that classification as the design brief, not as a threat: FLEX is architected against the high-risk requirements — risk management, data governance, technical documentation, record-keeping, transparency, human oversight, accuracy and robustness.
FLEX recommends; your operators decide. There is no autonomous switching path in the product. Recommendations arrive ranked, with expected effect and margin.
AC power flow, N-1 security and regulatory constraints are enforced inside the search. An action that violates them is never proposed — not proposed and flagged.
Every recommendation carries its reasoning: which constraint it relieves, by how much, and what the twin predicts if you do nothing instead.
Before anything reaches operations, FLEX is validated against your own historical power flows — the system must prove itself on situations your grid has actually seen.
Every recommendation, input state and operator decision is recorded — the record-keeping Annex III expects, and the evidence base for your own review.
EU hosting or fully on-premises inside your perimeter, air-gapped if required. Your grid model never trains anyone else's system.
On human-AI interaction in safety-critical infrastructure, see our peer-reviewed iScience (2025) publication — linked on the publications page.
There is no autonomous switching path in the product. Recommendations arrive ranked, with expected effect and margin.
Trust develops through phased deployment, allowing humans and AI to learn from each other over time. Every AI recommendation comes with clear reasoning, confidence scores and human override capability — and operators retain final decision authority at every phase.
Digital twin fundamentals, AI decision-making basics, override procedures and emergency drills — practised in a safe simulation environment. Typically 2–4 weeks.
AI suggests, humans decide. Operators compare AI against their own decisions and validate performance — trust built through transparency. Typically 1–2 months.
AI actively recommends actions; operators approve, modify or reject, with full override capability maintained and continuous performance monitoring. Full trust typically develops in 3–6 months.
Observed with operators using the digital twin in production settings: ~92% recommendation acceptance rate — with 100% human override capability retained.
