AI in the power grid – where should we draw the line? Kopier kobling

For the energy sector, this is an interesting backdrop. We have vast amounts of data, complex systems and many processes where AI can provide better analysis, more automation and faster decision-making. At the same time, grid operators manage critical infrastructure. This does not mean we should keep AI at a distance, but we do need to think carefully about how we use it.

The key question is therefore not whether we should use AI, but what we should use it for, what data it should have access to, and how much authority we are prepared to give it.

An AI assistant summarising a document is one thing. A model identifying abnormal voltage conditions in the low-voltage grid is something else. And a system that can actually make changes in an operational system is something else entirely.

In practice, we will probably use several types of AI at the same time. Generative AI can support search, documentation and assistants. Machine learning can be used for forecasting, anomaly detection, energy losses and grid capacity analysis. Increasingly, we will also see AI agents that can retrieve information from multiple systems and carry out parts of a workflow.

This is where the discussion becomes particularly interesting. There is a significant difference between a system saying “This meter should be investigated” and saying “I have disconnected the meter.”

The EU AI Act takes a risk-based approach. Critical infrastructure is one of the areas where certain AI systems can be classified as high-risk. This includes AI systems intended to be used as safety components in the management and operation of electricity supply.

But this does not mean that all AI used by a grid operator is automatically high-risk.

A forecasting model, a data-quality tool or an internal assistant will not automatically fall into this category. The intended purpose of the system and how it is actually used are important. The European Commission’s guidance gives examples such as network-load prediction and technical support tools as uses that can fall outside the critical-infrastructure high-risk category.

For grid operators, it is therefore useful to ask a few simple questions:

• What data can the AI access?

• What decisions can it influence?

• What is it actually allowed to do?

• What happens if it gets something wrong?

A useful distinction going forward is between data, decision and action.

At the first level, AI can identify errors in meter data, detect unusual events or find patterns that would be difficult for people to spot. At the next level, AI can recommend what the grid operator should do. The final step is allowing AI to carry out the action itself.

The closer we get to that final step, the more important control, logging, security and human oversight become. This also fits well with the requirements in the AI Act for high-risk systems, including risk management, traceability, human oversight, robustness and cybersecurity. Under the current timetable, the rules for high-risk systems in areas such as critical infrastructure will apply from 2 December 2027.

This does not mean that a person needs to approve everything an AI system does. That would remove much of the benefit. But we should be clear about where the boundary for autonomy lies.

An AI agent could, for example, be allowed to retrieve data, perform analysis, compare historical information and recommend an action. A critical operational command might still require explicit approval, while less critical tasks could be fully automated.

Much of the AI discussion is about models. For the energy sector, I believe the data will be at least as important.

Grid operators already have enormous amounts of data from smart meters, substations, sensors, customer systems and other operational systems. The challenge is often not a lack of data, but understanding what the data means, whether it can be trusted and how different data sources relate to each other.

A meter value on its own does not tell us very much. It becomes interesting when we know the metering point, the grid topology, the history and the other relevant context around it.

Good data management therefore becomes even more important with AI. Data needs to be accessible, quality-assured and understandable. At the same time, there must be control over what data AI can access and how it can be used.

It is also easy to equate AI with large language models. Many of the most interesting applications in the power grid will probably combine traditional machine learning, statistical models and generative AI. One model might identify an unusual pattern. A language model can retrieve relevant information and explain what has happened. An agent can then recommend or initiate the appropriate workflow.

It is when these technologies are combined with good energy data and existing operational systems that things become really interesting.

The AI Act and the role of critical infrastructure could easily make the energy sector more cautious than necessary. That would be unfortunate. The European Commission itself points out that the vast majority of AI systems currently used in the EU fall into the minimal- or no-risk category, and the regulation is designed around different requirements for different levels of risk.

There are already many areas where AI can create value without giving it control over critical operational processes: data quality, analytics, forecasting, documentation, customer service, troubleshooting and decision support are just some examples. From there, we can gradually move further.

For grid operators, the goal should therefore be neither “AI everywhere” nor keeping AI at a safe distance. It should be to use AI where it creates real value, while remaining in control of the data, the decisions and the actions.