Utilities run some of the strongest safety cultures in industry, then train for AI agents as if it were an IT cutover. Training ends at go-live, meter data is treated as beyond question, field teams are left out, and complaint responses drafted by the agent go out with commitments nobody checked. Their own safety habits would fix most of it.
Why is IT-led training the wrong model?
Utilities introduce agents through large change programmes, often run by IT with a fixed go-live. Training follows the programme's logic: system walkthroughs in the weeks before cutover, a support period after, then the programme closes.
Agent oversight does not end at go-live. An agent's certificate lapses after 90 days, a run that clears no depth drops it to depth 1 the same day, and capabilities widen as evidence builds. Each change alters what people see in the held queue. A training plan tied to a programme closure date leaves the business with reviewers trained for the agent as it was at launch. Move ownership of training to the operational teams that own the processes, and give it a standing budget rather than a project line.
Why isn't meter data beyond question?
A read from the utility's own smart metering system is sealed: issued by the system of record and tamper-evident. Staff often take that to mean it is correct. It means the read came from the meter; it does not mean the meter is working.
Faulty meters, crossed meters and communication errors produce sealed reads that are wrong. A billing agent may draft a large catch-up bill entirely on sealed data, and a reviewer who has learned that sealed means safe will approve it. Teach the difference between provenance and accuracy: provenance tells you who stands behind a fact, and judgement still asks whether it makes sense. A read that implies a household used five times its normal consumption deserves a second look, whatever its class.
Who gets left out in the field?
Utilities train customer operations and billing teams, because that is where agents first appear. Field operations come later, if at all.
But agents increasingly schedule work. A scheduling agent assigns routine jobs within the utility's own limits at depth 3 and holds anything needing a permit, a safety assessment or a shutdown. Field supervisors and planners decide those held jobs. Contractors who carry out metering and maintenance work often update the records the agent relies on. If supervisors have not been trained to read why a job was held, and contractors do not know that their job notes become evidence, the agent's decisions rest on weak ground. Include field supervisors and key contractor staff from the start, using jobs from their own region.
What goes wrong with complaint responses?
Utilities handle many complaints, and agents can draft responses from the account history, the contact record and the utility's policies at depth 2. The drafts are polite and complete, and handlers under pressure approve them.
The danger is in the commitments. A draft that promises a goodwill payment, a bill correction or a call-back by a set date binds the utility. Handlers must check every commitment against what the utility can and will do, and against the customer's actual history, before sending. They must also spot when a complaint reveals vulnerability or a safety issue that changes the response entirely. Train on real past complaints, including ones that escalated because of an unkept promise.
How can safety culture help?
Utilities already know how to build habits around risk: near-miss reporting, toolbox talks, stop-work authority, learning reviews. Agent oversight fits these habits well, and borrowing them costs little.
Treat an approval that should have been refused as a near miss and report it without blame. Run short toolbox-style talks on a held case of the week. Give every reviewer explicit authority to refuse or escalate, the equivalent of stop-work, and back them when they use it. Utilities that are essential entities under NIS2 in the EU also carry incident reporting duties, so make sure reviewers know how to escalate an agent behaving unexpectedly. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What should a utility do instead?
- Give agent training a standing owner and budget in operations, not a programme closure date.
- Teach the difference between provenance and accuracy, using meter cases.
- Include field supervisors and contractor staff who touch agent decisions.
- Train complaint handlers to check every commitment in a draft.
- Borrow safety habits: near-miss reporting, short regular talks, authority to refuse.
A utility can adopt the practice of tying approval rights to a current certificate, depth by depth. The general order is in the guide to training teams for AI agents. For detail, see what an AI agent training programme for utilities should include, the AI agent training plan template for utilities, and how much AI agent training costs for utilities. Insurers make a similar mistake with partner staff, covered in what insurers get wrong about AI agent training.
Where the Academy fits
Pryme Intelligence Academy's Operator track includes modules on keeping knowledge current and on when something goes wrong, which match the habits above. Certification requires a written exam and two weeks of real review decisions with no missed red line, and it is valid for 12 months, with renewal through an update module and a short quiz.
Questions people ask
Should contractors be certified as reviewers?
Only if they approve held actions in the utility's workspace. Most contractors need a short briefing on how their records feed the agent and why accuracy matters. Hold any contractor who approves to the same standard as staff.
How do we train control room staff?
If agents draft outage notices or schedule work that control rooms approve, control room staff need reviewer training on those specific actions. Keep agents away from real-time operational control; the training should focus on the drafts and schedules they do touch.
What is a good first case for a utility's training?
A catch-up bill built on a sealed but faulty meter read. It teaches provenance, judgement and the customer impact of a wrong approval in one example most billing teams will recognise.