The uncomfortable question facing utility operators deploying AI at the edge

Battery storage fleets running predictive models at the site level. Renewable asset managers using machine learning to schedule maintenance across installations scattered across a country. Grid-connected systems starting to lean on edge-based inference for decisions that cannot wait for a round trip to the cloud. None of this is hypothetical anymore. European utility operators are doing all of it, and the pace is picking up.

This is good. The use cases are credible, and the pressure to wring more value from distributed energy resources is only going to grow. But there is an uncomfortable question underneath all this activity that does not get nearly enough attention: once you push an AI model out to an edge node that is embedded in live energy infrastructure, how do you actually know it is still doing what you designed it to do?

Not on day one. Day one is easy. Everything has just been configured and tested, the data is clean, the parameters are well defined. I mean, six months later, after conditions have shifted, after firmware has been updated on some sites but not others, after a new OEM's equipment got added to the fleet with a different data profile. That is when things get interesting, and not in a good way.

Models do not age gracefully in the field

The moment a model goes live in a distributed energy environment, the world it was trained on starts to change around it. Assets from different manufacturers produce data in different formats with different naming conventions. Service agreements get renegotiated, altering what data is accessible. New equipment shows up with protocols that did not exist when the model was built. Seasonal and market conditions shift the patterns the model learned from.

All of this introduces drift. Not the kind that announces itself with a failure, but the slow, quiet kind. An optimisation recommendation that is slightly off. A diagnostic that misses a pattern it would have flagged six months ago. A predictive maintenance alert that fires too often or not often enough. Each one on its own looks like noise. Together, they erode the trust operators need to actually rely on these systems for real decisions.

In a centralised cloud setup, you can monitor for this. In a distributed environment with hundreds or thousands of edge nodes, many of which operate in disconnected or lights-out conditions, drift can go unnoticed for months. That is the governance gap, and it is getting wider as deployment accelerates.

The industry is right to move carefully

I have heard people describe the energy sector's cautious stance on AI autonomy as the industry "falling behind." That framing is wrong.

Energy systems are safety-critical infrastructure. When AI influences decisions about charge and discharge cycles in a battery storage system, or when it feeds into control logic for a grid-connected asset, the consequences of getting it wrong are physical. Not a bad product recommendation. Not a glitchy user experience. Actual equipment, actual grid stability, actual safety implications.

The consensus across the sector right now, and I saw this reinforced recently in conversations with operators from EDF, ABB, and Fluence, is that AI should sit in an advisory role with humans in the loop. That is not timidity. It is an appropriate response given where both the technology and the governance frameworks around it currently stand.

What the industry needs is not a faster path to autonomy. It needs better infrastructure for managing AI responsibly once it is deployed. And that means treating governance as something you build into the edge platform, not something you figure out after the first model misbehaves.

Have you read?
AI energy consumption and data centre rating under spotlight
Saudi Energy and Kraken Technologies form joint venture

Four things operators actually need

Having worked with utility operators and energy infrastructure providers deploying edge software across distributed environments, I keep coming back to the same four requirements.

Lifecycle orchestration. Models will need to be updated, patched, and sometimes rolled back. Energy assets are built for 20- to 25-year lifetimes. The AI running alongside them will go through many iterations over that span. You need robust mechanisms to push updates across hundreds of nodes, confirm they landed correctly, and revert to a known-good state when they did not. In air-gapped or intermittently connected environments, which are common in distributed energy, this is harder than it sounds.

Continuous observability. You cannot govern what you cannot see. Operators need fleet-wide visibility into how models are performing over time, not just at the point of deployment. That requires telemetry and logging designed for environments with unreliable connectivity and varying hardware configurations from site to site. Without this, drift stays invisible until it surfaces as an operational problem.

Explainability. This one is becoming non-negotiable, particularly in Europe. The EU AI Act is establishing new requirements for high-risk AI systems, and energy infrastructure is likely to fall within that classification for certain use cases. But even setting regulation aside, operators who cannot explain why a model recommended a particular action will not trust that recommendation. And they should not. In environments where a bad decision has physical consequences, black-box outputs are not acceptable.

Clear boundary definition. What is AI permitted to advise on? What can it act on with oversight? What stays exclusively under human control? These boundaries need to be defined and enforced consistently across the fleet, not left to individual site configurations or informal practices within teams. With agentic AI approaches starting to emerge, getting these boundaries right now matters more than it did even a year ago.

European operators face a layer of complexity that their counterparts elsewhere do not. The EU AI Act is progressively entering into force, and it will impose specific obligations on operators running AI systems classified as high-risk. The practical details are still being worked through across the industry, but the direction is unmistakable: you will need to demonstrate that you have processes for monitoring AI performance, managing risk, and ensuring transparency.

Organisations that treat this as a retrofit exercise, bolting governance onto existing deployments when regulatory deadlines arrive, are going to struggle. The ones building these capabilities into their edge infrastructure now will be in a far stronger position.

Where the conversation needs to go

For several years, the dominant question in this space has been about where AI can be deployed in energy operations and what value it can deliver. That question has been largely answered. The technology works. The use cases are proven.

The harder question, and the one that will determine whether edge AI becomes a trusted long-term operational capability or stalls at pilot scale, is about governance. How do you maintain model reliability over time? How do you manage updates and rollbacks across a distributed fleet? How do you keep humans appropriately in the loop while still capturing the speed and efficiency that AI promises?

These are not glamorous questions. They do not generate the same excitement as a new model architecture or a breakthrough in inference speed. But they are the questions that separate organisations deploying AI from organisations that can actually depend on it. And in energy, dependability is the whole point.

About the Author

Andrew Foster is the Chief Product Officer at IOTech, with over 20 years of experience developing IoT and Distributed Real-time and Embedded (DRE) software products. He has held senior roles in Product Delivery, Management, and Marketing, and frequently speaks at industry conferences on distributed computing, middleware, embedded technologies, and IoT.