Is an energy management system an "AI system"? It depends on what it actually does
The "AI" label now appears on every energy management offer. With national supervisory authorities able to run inspections and impose penalties for AI Act breaches, the question has stopped being a marketing one. Except that the answer is neither yes nor no.
The question customers are asking
After the autumn attacks on refrigeration controllers in European stores, the question keeps coming back: are remote control and "intelligent" systems a risk in themselves? The first wave ran through around 700 stores in France on the night of 11 October; the second, on the night of 1 November, reached more than a hundred sites in further countries.
That is a separate subject with separate accountability — we covered it in the piece on controllers left on the open internet. What follows is the other half of the same question: what are we actually buying when the word "AI" appears in an offer, and what obligations come with it?
Not every optimisation system is an AI system
The European Commission issued guidelines on the definition of an artificial intelligence system — document C(2025) 5053 final of 29 July 2025. They are not binding; only the Court of Justice can give an authoritative interpretation.
The guidelines name four categories of solution that fall outside the definition: systems for improving mathematical optimisation, basic data processing, systems based on classical heuristics, and simple prediction systems.
The first category matters most in this industry. According to the guidelines, systems used to improve mathematical optimisation, or to accelerate and approximate traditional, well-established optimisation methods, fall outside the scope of the definition — because, while such models have the capacity to infer, they do not transcend "basic data processing".
The caveat matters, though. The guidelines say such systems may be excluded, and only where their self-adjustment serves computational performance rather than adjusting decision-making models in an intelligent way.
Energy appears on both sides of the guidelines
This is where the convenient answer — the one vendors like to quote half of — runs out.
Among the examples outside the definition, the guidelines list machine learning models that accelerate physical simulations, and a satellite telecommunications system optimising the allocation of power and bandwidth. But in the part describing systems that are covered, an example lands squarely in our own field: AI systems for energy consumption, designed to estimate consumption by analysing data from smart meters, weather forecasts and behavioural patterns.
The guidelines close the matter with a sentence worth remembering: no automatic determination, and no exhaustive list of systems inside or outside the definition, is possible.
So neither "it is only optimisation, so this does not concern us" nor "it is AI, so we are in the strictest regime" is an answer. The answer is a description of what the system does.
High risk is about function, not sector
The second misunderstanding concerns the high-risk category. Using AI in energy does not place a system there automatically.
Annex III to the AI Act, in its critical infrastructure point, covers AI systems intended to be used as safety components in the management and operation of critical digital infrastructure, road traffic, or in the supply of water, gas, heating or electricity. What decides is the safety-component function, not the fact that the work happens in energy.
Regulation (EU) 2026/1744 of 8 July 2026, in force since 27 July 2026, amended the AI Act among others. Legal commentary indicates that it narrowed the definition of a safety component: a system whose failure does not endanger health or safety is not a safety component. Before relying on that in a specific project, check the text of the regulation itself.
These are not the only rules in play
Alongside the AI Act sit cybersecurity requirements, and their application depends on the entity and the product.
The Cyber Resilience Act (Regulation (EU) 2024/2847) introduces reporting obligations from 11 September 2026 — covering actively exploited vulnerabilities and severe incidents — with the main product obligations applying from 11 December 2027. National cybersecurity legislation adds its own requirements on top.
For a buyer the practical conclusion is simple: classifying a solution is not one box but several, and the vendor should be able to fill in each of them.
Three questions for the vendor
Instead of asking whether a system "has AI", ask something specific.
- What does the AI actually do — describe a chart, forecast, or determine setpoints? That distinction decides the classification more than anything in the marketing material.
- Have the limits of possible decisions been clearly defined — what are the setpoint boundaries, the process requirements and the comfort thresholds, and who approves them?
- How is access to the automation organised, and who is accountable for maintaining it — what belongs to the vendor, what to the integrator, what to the user, and how is that checked across the whole service life?
Energy is saved by equipment running at a different setpoint or on a different schedule. That takes measurement, control and verification of the effect. The vendor should also be able to explain how their solution is classified and what follows from it.
The name of the system is not a substitute for any of that.
What decides how an energy management system is classified under the AI Act is not the label on the offer or the sector it serves, but what the system actually does and whether it acts as a safety component — and the Commission's guidelines give examples on both sides of the definition.
Percee is an EMOS-class layer that measures, decides, acts and verifies the result within clearly defined limits: setpoint boundaries, process requirements and comfort thresholds are set before control is switched on, and every change is recorded together with its reason. The system is segregated all the way down to the level at which it talks to the automation, and it passes the cybersecurity audits of demanding customers, including banking and industrial clients. Questions about classification, decision scope and maintenance responsibility can therefore be answered with a document rather than an assurance.
See how Percee works →