AI in energy and utilities earns its keep in the paperwork and triage that surround the network, not in the control room. The strongest operator-level use cases are reading and reconciling field work orders and permits-to-work, keeping asset and maintenance records consistent, clearing meter and billing exceptions, triaging outage reports and complaints, assembling regulatory filings, and checking the contractor invoice-and-GRN chain — while switching decisions, isolation, and anything that puts a person near energised plant stays entirely human. Utilities are document businesses wrapped around physical infrastructure, and the documents are where the leakage hides.
That framing matters because the sector is routinely sold the opposite: autonomous grid optimisation, self-healing networks, prediction of everything. Some of that is real research; little of it is where a distribution utility, a generation company, or a city water business loses money this quarter. The reliable pattern is the one that holds across AI in operations generally: the machine reads, extracts, and flags; the person decides. In a sector where the wrong decision electrocutes someone, that boundary is not a preference — it is the licence to operate.
What does a utility's document load actually look like?
Before the use cases, the inventory. A mid-sized utility runs on a paper chain most outsiders never see:
| Workflow | The documents | What goes wrong manually |
|---|---|---|
| Field work | Work orders, job cards, completion reports, site photos | Illegible closures, work recorded against the wrong asset |
| Safety | Permits-to-work, isolation certificates, method statements | Mismatched isolations, expired authorisations found late |
| Assets | Maintenance histories, test records, nameplate data, drawings | Histories split across systems, records that stop mid-life |
| Revenue | Meter reads, billing exceptions, disputed bills | Exceptions queue for weeks; estimates compound |
| Customers | Outage reports, complaints, no-supply calls | Duplicate tickets, misrouted faults, slow pattern detection |
| Regulation | Licence returns, incident reports, tariff filings | Manual compilation from a dozen sources under deadline |
| Contractors | Invoices, GRNs, measurement sheets, rate schedules | Invoices paid against unverified quantities |
Every row is a reading-and-reconciliation problem at volume. That is the shape AI handles well.
How does AI help with field work orders and permits-to-work?
Field work generates unstructured returns: a completion report typed or scrawled by a crew at the end of a shift, photos, a job card that may or may not name the right asset. AI can read those returns, extract what was done, to which asset, with which materials, and post a structured closure — flagging the ones where the record contradicts the order, the asset ID does not exist, or the photos show something the text does not say.
Permits-to-work deserve their own sentence, because they are where documentation and safety meet. A permit system's failure modes are documentary: an isolation listed on the permit that does not match the switching schedule, an authorisation signed by someone whose competency has lapsed, a permit still open against plant about to be re-energised. AI can cross-check these documents against each other and against the asset register, and flag mismatches for the authorising engineer before work starts. It prepares and checks. The engineer authorises. No credible design puts a model in that signature line.
What can AI do with asset and maintenance records?
The asset base of a utility outlives every system that has ever recorded it. Histories for a forty-year-old transformer sit partly in an EAM system, partly in scanned test sheets, partly in a retired engineer's filing habits. AI is genuinely good at the archaeology: reading scanned test records and nameplate photos, extracting dates, readings, and serial numbers, and consolidating a per-asset history that maintenance planning can actually use.
The honest caveat is the same one that applies everywhere: consolidation is not prediction. Predictive maintenance needs meaningful failure history and clean condition data, which many utilities do not yet have in usable form. Building the record first is not a consolation prize — it is the prerequisite, and it pays for itself in avoided repeat inspections and better-informed replacement decisions even if no model ever predicts anything.
How does AI clear meter and billing exceptions?
Revenue leakage in a utility concentrates in the exception queue: reads that fail validation, consumption that jumps implausibly, meters recorded against the wrong premises, estimated bills compounding for months, disputed invoices sitting with a back office. Each exception is a small investigation — compare the read history, the meter record, the tariff, the site notes — and the investigation is mostly reading.
AI can do that first pass at volume: assemble the relevant history for each exception, classify the likely cause, resolve the mechanically obvious ones under rules, and hand the genuinely ambiguous ones to a person with the evidence already gathered. The economics mirror the general case for removing manual data entry: per-item minutes, multiplied by volume, plus the error cost of tired humans keying under backlog pressure.
Can AI triage outages and complaints?
Yes, at the intake layer. No-supply calls, complaint emails, and social reports arrive as unstructured text across channels, frequently duplicated. AI can classify them, extract location and symptom, deduplicate against known incidents, and route: probable network fault to the control room queue, billing dispute to revenue, recurring complaint at one address flagged as a pattern. Faster triage shortens the time between the first report and someone competent seeing it, which is often the largest controllable slice of a response time. Dispatch itself — which crew, which priority, whether to switch — remains a control-room judgment.
Regulatory filings and the contractor chain
Two quieter workflows round out the picture. Regulatory returns — licence compliance, incident reporting, performance statistics — are compilation exercises across many internal sources against fixed deadlines; AI can assemble the draft and cite where each figure came from, with the regulatory affairs team owning what is submitted. And the contractor chain — invoices against measurement sheets against GRNs against rate schedules — is a three-way-match problem identical in shape to the ones in supply chain operations: extraction plus cross-checking, with mismatches flagged for a quantity surveyor rather than paid on trust.
One sector-specific note on ownership. Survey work on regional AI adoption (McKinsey's State of AI in the GCC, cited in our research) finds that in energy and utilities, central IT leadership is the most common owner of AI programmes. That has a practical implication: operations teams with a concrete document problem usually get further by bringing IT a measured workflow than by waiting for an enterprise AI strategy to reach them.
What stays human in utility operations
Plainly: switching and dispatch decisions; permit authorisation and isolation confirmation; any safety-critical judgment in the field; final sign-off on regulatory submissions; disconnection decisions affecting customers, which carry legal and human consequences no model should own. AI in a utility is a clerk and a checker with perfect stamina — never an operator.
Where to start
Pick the exception queue you can already count — billing exceptions and contractor invoices are the usual candidates, because the backlog is visible and the per-item cost is measurable. Baseline it honestly: items per month, minutes per item, error rate, ageing. Add AI as a first-pass reader with a person on every resolution that touches a customer or a payment, prove the numbers on that one queue, and only then move toward the harder territory of asset records and analytics. The utilities that get value from AI are not the most instrumented — they are the ones that were precise about which paperwork was drowning whom.