Agents are joining your O&M team

Walk onto any renewables operations office on a Monday and you’ll find skilled people doing work that doesn’t need them. Reconciling meter reads against settlement runs. Re-keying fault data from one maintenance system into another that won’t talk to it. Chasing a missing timestamp so a report can go out. These are engineers, analysts and traders, and a good chunk of their week goes on shuffling paper between systems that were never designed to connect.

That’s the part of the energy transition nobody puts on a conference slide. We’re building wind, solar and storage faster every year. Every MegaWatt added is another stream of operational admin behind it: more assets to monitor, more contracts to settle, more exceptions to investigate. The clean-energy build-out has an unglamorous operations bill attached, and so far we’ve mostly paid it by adding headcount.

There’s an alternative arriving, and it gives us back the thing humans are actually good at: judgement.

Agents, not dashboards

For fifteen years at EDF I worked across nuclear, trading, battery storage and renewables, and I watched the industry build a lot of dashboards. Dashboards tell you something is wrong. Somebody still has to read the dashboard, work out what to do, and go and do it. The reporting got better; the doing stayed manual.

What’s changing is that software can now take the action, not just flag it. An AI agent can read the fault alert, pull the asset history, draft the work order, check it against the maintenance schedule, and put it in front of a human for sign-off. The person still decides. The agent does the forty minutes of gathering and drafting that used to sit in front of the decision.

That’s the shift I’d put at the centre of any honest conversation about AI in energy: we’re moving from systems that report to systems that act, with a person in the loop. Human teams will work alongside agent teams the way we currently work alongside spreadsheets, except the agent does the tedious part and hands back the judgement call.

And that’s a good thing. The agent takes the crap work. The people get the interesting work back: the engineering judgement, the commercial calls, the edge cases where experience actually matters.

Where it pays off first

The macro version of this story, AI balancing the national grid in real time, gets the headlines. I’d start somewhere less dramatic and far more achievable: the micro-operations of individual fleets and portfolios.

A few examples from work I’ve been close to:

  • At Low Carbon, it started with an audit. Not the high-level strategy deck you get from a big-four firm, but a targeted look at the actual workflows and time sinks in the operations team. That found the real problem: Work Orders sitting in a backlog waiting for human review. The project that follows will triage this automatically, so the team can get moving with corrective actions. Most of the value isn’t a clever model, it’s the automation of some clear rules.

  • At Aukera, a developer moving into operations for the first time, the operational admin is all brand new. There’s no legacy ops team to inherit, which makes it the right moment to build the agent-augmented way of working in from the start rather than retrofitting it later.

  • At Calton and SEG, both running field-services operations, the same pattern showed up: scheduling and job admin spread across systems that don’t share data cleanly, so coordinators spend their mornings re-entering the same information in three places. Automating those handoffs is provable, you can count the hours that go back to the people doing the actual coordination. Field-services O&M admin is one of the clearest places to show this works.

Notice the pattern. None of these is science fiction. Each is a real operational process where the bottleneck is admin and integration, not intelligence. That’s where AI earns its keep in energy today: O&M admin, joining up systems that don’t speak to each other, and turning reporting into action.

The bit that makes it safe

Here’s the part that matters to anyone with a shareholder, and it’s the part I’d want a CTO to push hardest on.

Large Language Models work probabilistically. That’s a real difference from the IT we’re used to. A traditional system, given the same input, gives the same output every time. An AI agent is reasoning over messy, ambiguous data, and it can be wrong in ways a purely scripted rules engine can’t. If you bolt one into a settlement run or a dispatch decision and walk away, you’ve amplified risk, not removed it.

So you don’t walk away. You design the human-in-the-loop deliberately: the agent does the gathering, drafting and reconciling, and a person validates before anything irreversible happens. Done well, that’s a net reduction in risk, because the agent is also a tireless checker. It catches the missing read, the mismatched timestamp, the exception a tired analyst would skim past at 5pm on a Friday.

Reducing operational risk while freeing people up: that’s the line I’d lead with in any boardroom, and it’s true, as long as the controls are built in from the start rather than bolted on after.

What good looks like

The organisations getting value from this aren’t the ones with the biggest AI budgets. Plenty of energy companies are spending heavily right now without being able to name what they’re getting back. The ones pulling ahead are doing something more boring and more effective: picking a single painful operational process, automating the admin around it, keeping a person in the loop, and measuring the time and value they get back. Then they do the next one.

It compounds. Every process you take off your team’s plate is capacity you don’t have to hire for as the portfolio grows. Given how fast we’re adding renewable and storage assets, that matters. The transition needs us to operate far more generation with people we can’t recruit fast enough. Agents working alongside those teams are how the maths starts to work.

We’re going to be running the energy system with people and agents side by side. The agents will do the paper-shuffling. The people will do the work worth doing. For an industry that’s about to operate more assets than ever, that’s not a threat. It’s the only way the numbers add up.

Tom Reynolds spent fifteen years at EDF across nuclear, energy trading, battery storage and renewables before founding Flow Vendor Ltd, an automation consultancy for the energy sector. Flow Vendor helps asset owners and operators put AI to work in operations, with humans firmly in the loop. More at flowvendor.energy.

AI