If AI is the answer to the energy industry's problems, what is the question?

Five years ago, a data scientist joined a UK distribution network, waited two weeks for a laptop and was then told he couldn't install Python. Today, on the 8am into Cannon Street, he fixes a bug on his phone with an AI coding assistant. This tells you a great deal about how fast technology has moved but rather less about whether we are asking the right questions of it.

Earlier this year, Softwire and Energy UK brought together practitioners to discuss the impact and value of AI in the sector. What questions are they asking of technology, and are they getting the answers they hope for?

"If the only tool you have is a hammer, it is tempting to treat everything as if it were a nail," states the Law of the Instrument, attributed to Abraham Maslow. It describes a bias that suggests when people have a specific tool, they apply it to all situations regardless of its suitability.

Sam Young, AI practice manager at Energy Systems Catapult, believes that the temptation to turn to AI as a solution to all energy sector problems should be resisted: "If you solved this problem with AI, would anything change?” he asks. “No? Then don't solve it with AI, even if you can."

The principle sounds obvious. Yet it is surprisingly hard to apply when under pressure to be seen moving on AI, from suppliers promising efficiencies to leadership teams who have read the headlines and want a strategy to match. The consequence of getting it wrong is not only wasted budget and POCs that never leave the innovation ‘lab’, but also the loss of the valuable applications that could have instead been developed: the ones that could actually help a control room operator avoid a blackout or increase the capacity of Power Systems Engineers (as Sam Young put it: "Who has too many Power Systems Engineers? No one!").

So, rather than haphazardly swinging the hammer, ask which problems you have ignored because they seemed too hard, and whether advances in technology, including AI, now make them tractable.

Sam's favourite example of AI done well is HeatGeek's Zero Disrupt tool. The starting point was not "how do we use AI in heat pump installation?" but "what is slowing installers down?" The answer turned out to be a tangle of manual calculations, uncertainty about what to skip, and a process that asked too much of the engineer standing in front of a customer's boiler. Zero Disrupt broke that process apart and applied appropriate technology at each point of friction. The result was that installation time was halved and costs fell by 75%. It also changed perceptions; one AI sceptic told Sam he loved the product without quite registering the technology behind it. That is the standard worth aiming for, not "this is AI-powered," but: this works, and you barely notice why.

Jamie Bright, data science manager at UK Power Networks, offered another example he experienced in his team: a key engineer left, taking years of undocumented knowledge in the system with them. "It was like, oh my god, we've got this technical debt now," Jamie said. "It's now my responsibility." But using an AI tool, he was able to work through the inherited code, asking it where to look, explaining the process, and using it to understand what had been built. Now, the fear of that knowledge gap, he said, had been largely alleviated not because the problem went away, but because the right tool is now available.

Whilst public attention has been focussed on generative AI, the panel were clear that machine learning, generative AI, and agentic AI are not interchangeable. Each suits different problems, carries different risks, and requires a different level of oversight.

Classical predictive models, for example, tend to be considerably more transparent and auditable than their generative counterparts. And while both large language models and forecasting models can produce nonsensical outputs the hallucinations generated by a LLM are a different class of failure entirely from those from a predictive model producing an output outside its expected range – an error that is bounded, measurable, and far easier to detect and correct.

Treating them as the same tool can overstate risk in one area, misplace investment in another, and miss the fact that in an engineering-heavy sector, much of the highest-value opportunity still lies in machine learning.

Ruby Mitchell, who leads AI initiatives at Kraken, has spent a year working on AI enablement, her sessions focussing on questions seen across the sector: what it can do, how to use it, and which tools are approved? But over time, she found herself changing the angle and looking at ‘problem framing’ instead of directly addressing AI: what are your team's actual constraints, where do things slow down, and if you needed to move twice as fast, what would need to be true?

While this might be a small shift in how you run a workshop, it might end up significantly shifting what you end up building. Organisations that treat AI adoption primarily as a technology question tend to get individual experimentation and little beyond it. Ruby pointed out that "team adoption trumps individual use". The ones asking what problems actually need solving tend to build things their teams depend on.

The sector is going through one of the biggest changes since the electrification of the grid itself, the expertise to manage that change is scarce and getting scarcer, and the window to get this right matters not just commercially but for every customer who depends on the system working.

So, how do we know when and where to deploy AI to solve the sector's challenges? Here are three questions that any AI initiative in the sector should be able to answer before it scales: