A team from Tampere University spent four weeks embedded inside a mid-sized Nordic energy company, the kind that runs trading, heating, solar and wind under one roof and stitches together more than 100 platforms to serve thousands of customers. They ran 16 interviews across nine departments. They came out with 41 places where generative AI could plausibly help.

Zero of those 41 were running in production when the researchers left.

Thats not a story about a company that doesnt get it. A senior representative who oversees AI work coordinated the entire study and had real executive backing behind it. This is what serious, well-resourced AI adoption actually looks like from the inside, in a sector where a bad automated call touches the grid, the billing system, and the price customers pay.

Energy companies plan in decades, not quarters, and that shapes what they actually want a model to do.

41 Use Cases, Six Buckets, Three Real Priorities

The researchers coded everything into six categories: reporting, RAG-based retrieval, predictive maintenance, anomaly detection, budgeting and forecasting, and a leftover bucket of department-specific asks. When they ranked all 41 by business importance and ease of implementation, three things rose to the top across nearly every department. Automating reporting. Predictive maintenance to cut downtime. Better forecasting for long-range planning.

The forecasting appetite is sector-specific and its a big one. One participant described planning out to 2036, deciding how many customers the company will have and when overhead lines get swapped for underground cable. Another put it plainly: “First one is like a strategic forecast of changes in the operating environment in the electricity distribution network. So it is like forecasting thing. Different kind of things that we forecast. What is changing in our field here.” Energy companies think in decades. That reshapes what they actually need a model to be good at.

“We spend much time on Excel and manual reports, and often the same numbers are checked again and checked again.”
Finance department participant, quoted in the study

The Real Blocker Isnt The Model

Five themes came out of the interview coding, and forecasting wasnt even the biggest one. Data fragmentation and system integration logged 30 coded instances, nearly tying reporting and forecasting for first place. A participant from the systems and data integration team described the gap bluntly: “We have error handling, but many processes we do not have because the amount of work is so big. It could be better if some system monitored and categorized the errors automatically.” Another asked for the basics first: “Integration between systems, integrations for these other systems.”

This is the finding that should worry executives more than any capability gap in the model itself. The papers own conclusion is that the main issue isnt an absence of use cases, its the fit between new tools and how data already moves, or doesnt move, through the organization, which is exactly what an AI Assessment for companies is built to surface: the gap between where leadership thinks the data infrastructure is and where it actually is.

Compliance and validation work, invoice checks, vendor review, fraud detection, showed up as its own separate theme with 21 coded mentions. “Invoices often require corrections, which means extra work and double checking for compliance,” one Finance participant said. In an industry where a bad anomaly-detection call touches grid pricing and customer billing, that caution isnt hesitation. Its the job.

Every pilot the researchers built kept a human between the AI’s output and anything actually happening.

Two Pilots, Zero Autonomy Handed Over

The team built two working pilots to prove the concept wasnt just theoretical. The first was an email response system, built on LangChain and LangGraph, trained on the companys own historical email data. Measured against a reference dataset with BERTScore, it hit roughly 0.89, a strong semantic match to how a real employee would answer. The second was a document retrieval chatbot that lets staff search internal files in plain language instead of digging through folders across a distributed, cloud-based system.

Neither one sends anything without a human clicking approve. Thats not a technical limitation, its the design. Every generated email gets reviewed before it goes out. Participants said as much themselves. One wanted AI to flag issues rather than fix them, “or it would be easier if system assisted.” Another, describing a shrinking headcount, said the goal was reducing anything manual or time-consuming, not removing people from the loop entirely: “There will not be more people to do this, maybe less. So anything manual or time consuming that we can reduce helps the workload.”

Forty-one use cases mapped across nine departments is real progress. A held-out human approval on every single one of them, four weeks in, is the actual state of generative AI in the energy sector right now.