AI in Energy and Utilities: Use Cases Beyond the Control Room
Most conversations about AI in energy and utilities go straight to the control room. Load forecasting, grid balancing, predictive maintenance on turbines. Those are real problems, but they are also the best instrumented, most heavily invested corners of the business. The pain your teams complain about every week lives somewhere else. Field reports that arrive as photos of handwriting and never make it into the asset system. Compliance filings assembled by hand from spreadsheets, emails and half-remembered precedent. Outage communications written under pressure at 2am, inconsistent across channels. Asset records that drift so far from physical reality that crews stop trusting them. This is document-heavy, deadline-driven, judgement-laden work, and it is exactly where production AI agents now do their best work. Below are the use cases we see actually shipping inside utilities, with enough detail on how each one works for you to judge whether it fits your operation.
Field reports that file themselves
A crew finishes a job and the record of it exists as a voice note, a few photos and a form filled in from the cab of a truck. Someone in the back office retypes it days later, or nobody does.
An agent here reads the raw inputs: the transcribed voice note, the photos, the work order the job was raised against, and the asset's existing record. It decides what the report actually says. Which asset was worked on, what was found, what was replaced, whether the stated condition matches the photos, and whether anything described triggers a follow-up, such as a defect that needs a new work order. It writes back a structured completion record in your work management system, in your codes and your fields, and raises the follow-up orders it identified. It hands to a person whenever the evidence disagrees with itself: the photo shows a different transformer serial than the form, the described work does not match the order, or a safety-relevant defect is mentioned. Those land in a supervisor's queue with the discrepancy spelled out, not buried.
Compliance filings drafted from evidence, not memory
Regulatory reporting in this sector is relentless. Reliability standards, environmental returns, incident notifications, price control submissions. Each one means pulling data from several systems, checking it against the rules for that filing, and writing prose a regulator will read closely.
An agent reads the filing requirements you have codified, the operational data from the source systems, and your previous submissions of the same type. It decides which figures belong in which section, whether this period's numbers deviate from history in ways that need explanation, and where the data is incomplete. It writes back a full draft filing with every figure traced to its source, plus a list of gaps and anomalies. The handover point is fixed and non-negotiable: a named compliance owner reviews and signs every filing before anything leaves the organisation. The agent's job is to make that review take an afternoon instead of three weeks, and to make it easier, because every claim in the draft carries its provenance.
Outage communications that keep pace with the event
During a major outage the people who know what is happening are busy restoring supply, and the people writing customer updates are guessing. The result is generic messaging that call centres then have to apologise for.
An agent reads the live outage feed from your management system, crew dispatch status, and your approved messaging templates and tone rules. It decides which affected areas need an update, what can honestly be said about restoration times given current crew status, and which channel each audience needs: web banner, SMS, call centre script, or a briefing for the media team. It writes back drafted updates for each channel, consistent with each other and with the operational facts. For routine events, updates can flow on a short automatic timer after a duty manager approves the first message. For anything involving injury, extended restoration, or media attention, every message is held for human sign-off, and the agent's role narrows to keeping the facts assembled so the person approving is never working from stale information.
Asset records that stop drifting from reality
Every utility has an asset register, and every utility knows it is partly fiction. Nameplates disagree with records, replaced equipment keeps its predecessor's data, and location fields point at the wrong site.
An agent works through the register continuously rather than in a heroic one-off cleanse. It reads the asset record, then everything that mentions that asset: inspection reports, work orders, invoices, commissioning documents, photos from field visits. It decides where the record and the evidence disagree, and how confident that disagreement is. For high-confidence corrections backed by primary evidence, such as a serial number visible in a recent photo, it writes the correction back with a full citation trail, if your governance allows automated writes at all. Many utilities sensibly start with the agent proposing every change. Anything ambiguous, and anything touching criticality ratings or safety classifications, goes to an asset engineer as a proposed change with the evidence attached. You get a register that converges on the truth instead of decaying from it.
Connection requests and the queue behind them
New connections, solar interconnections, EV charging installations. The queue grows faster than the team assessing it, and most applications stall on missing or inconsistent information rather than hard engineering questions.
An agent reads each incoming application, the attached single-line diagrams and specifications, and your connection standards. It decides whether the application is complete, whether the declared equipment matches the diagrams, and whether the request fits within thresholds your engineers have defined for straightforward cases. It writes back a completeness assessment, a request to the applicant listing exactly what is missing, and a pre-populated technical summary for the engineer who will assess it. The engineering judgement itself stays human. Every approval decision is made by a qualified person, but that person now opens a file that is complete, consistent and summarised, instead of spending the first hour discovering what is absent.
Contractor work verified against what was agreed
A large share of network work is delivered by contractors, and verifying their invoices against contracted rates and evidence of completion is slow enough that much of it gets sampled rather than checked.
An agent reads the invoice, the purchase order, the contracted schedule of rates, and the completion evidence: sign-off forms, photos, timesheets. It decides line by line whether the billed items match the agreed rates, whether the claimed quantities are plausible against the work order, and whether completion evidence exists for what is being charged. Clean invoices are written back as approved-for-payment recommendations with the checks attached. Anything that fails a check goes to your commercial team with the specific mismatch identified: this line bills eight hours against a job signed off in three. Disputes with contractors remain a human conversation. The agent's contribution is that the conversation starts from evidence.
Where to start
Resist the instinct to start with the most valuable process. Start with the one that has three properties: the inputs are already digital or easily captured, the output has a clear owner who reviews it anyway, and an error caught at review costs annoyance rather than harm. Field report structuring and invoice verification usually qualify. Outage communications and compliance filings come next, once your teams have learned to supervise agents and you have learned where they fail. Run the agent in shadow mode first, producing drafts alongside the existing process, and measure the correction rate before anything it writes becomes the record. Give it a scoreboard from day one: items processed, items escalated, corrections made by reviewers. If the correction rate does not fall month on month, stop and find out why. And put real effort into the escalation paths, because the difference between an agent your supervisors trust and one they route around is entirely in how it behaves when it is unsure.
Where AI fails here
Honesty about limits matters more in this sector than most. Agents fail on inputs that were never captured: if the field crew took no photo, no system will conjure the asset's condition. They fail quietly on plausible-looking extraction errors, a misread serial number or a transposed quantity, which is why every automated write needs provenance and sampling, not blind trust. They cannot own regulatory accountability. A regulator holds your named responsible person to account, and no filing should leave without that person's review. They are the wrong tool for real-time protection and control decisions, which belong to engineered, deterministic systems with certified behaviour. They struggle where your own rules are undocumented, because an agent can apply your connection standards but cannot invent them, and the exercise of writing rules down is often the hardest part of the project. And they will not fix a broken process. Automating the production of a report nobody reads produces unread reports faster.
None of this requires you to recognise yourself in every use case above. It requires one process where the paperwork burden is real, the review step already exists, and your people are tired of being the integration layer between systems. That is where a production agent, built inside your environment and supervised by your team, pays for itself first and teaches your organisation how to run the next one. If a specific process and its specific mess came to mind while reading, we would genuinely like to hear you describe it to Adopt AI Partners.