Scale in a utility is measured in accounts, and every account the agent cannot settle alone becomes a held case for a person. Billing exceptions, payment plans, refunds and vulnerable customers all route there. With untrained reviewers, the held queue grows faster than the agent saves work, and the agent never earns the depth that would make it worth running.
Where does a utility's agent hand work to people?
The work agents do well in a utility is the high-volume back office: validating meter reads, explaining bills, setting up payment arrangements and handling credit balances. Each comes with a natural stopping point.
A billing agent reads meter data and flags reads that fall outside the expected range at depth 1. It drafts a corrected bill or a catch-up bill explanation at depth 2, sent by a billing analyst. A collections agent agrees payment plans within limits the utility has set at depth 3, so a plan within the customer's affordability band goes through and one outside it is held. A refund of a credit balance moves money, so it sits at depth 4 and waits for a named person. Outage notices are drafted at depth 2 and released by the incident team.
Some decisions stay with people whatever the agent's certificate says. Anything that could cut off or restrict supply, or change how a customer in a vulnerable situation is treated, belongs to a trained person with the full picture.
Why does the queue decide whether agents scale?
Utility work arrives in surges. A run of estimated reads, a tariff change, a cold snap or a storm can multiply exceptions overnight. The agent handles the routine cases without slowing down. The held cases pile up in front of people.
If those people are trained, a surge means a busy week. If they are not, it means one of two failures. Reviewers under pressure approve catch-up bills without checking the reads behind them, and complaints follow weeks later. Or they send every draft back for manual handling, and the agent adds a step instead of removing one. Either way the business learns that agents "do not work at scale", when the constraint was the people at the stopping points.
Telecoms operators face the same arithmetic with billing disputes and credits; the parallel is drawn in telecoms will not scale AI agents without trained reviewers.
How do reviewers affect the agent's depth?
Depth is earned, not configured. A utility may want its collections agent to agree payment plans at depth 3 across all customer segments. Attest decides whether the evidence supports that, by replaying the agent against real cases from the utility's own operation and scoring every answer against a floor. Those real cases include the decisions reviewers made on the agent's drafts.
When reviewers approve plans that breach affordability rules, or refuse sound plans out of caution, the record stops being a reliable guide to what good looks like. The agent's depth stays where it is. And because certificates lapse after 90 days, with a run that clears no depth dropping the agent to depth 1 the same day, a weak review culture shows up again at every renewal.
What must utility reviewers be able to do?
The core skills are the same as anywhere; the cases are not.
- Check the source behind each number. A smart meter read from the utility's own metering system is sealed. A read submitted by a customer is declared. A photo of a meter emailed in is external until checked, and it cannot authorise a refund.
- Spot vulnerability signals the agent has flagged or missed. A customer mentioning a medical device, a new baby or a recent bereavement changes what is appropriate, and the reviewer must act on it.
- Decide held payment plans within policy. Know the affordability rules, the limits and when to escalate to a specialist team.
- Refuse with a reason. Write why a catch-up bill or a refund was stopped, so the next reviewer and the agent's owner can learn from it.
- Hold the red lines. For example, never approve a refund to bank details changed in the same contact, and never agree a plan that leaves a vulnerable customer without supply.
Public bodies running services face a similar mix of volume and duty of care; public sector will not scale AI agents without trained reviewers makes the case for them.
Can a utility start before its reviewers are trained?
At low depth, yes. An agent that reads meter data and drafts explanations for staff to send is safe to run while reviewers learn on its drafts, and the shadow period becomes the training. The detail is in whether utilities can use AI agents before staff are trained.
What a utility should not do is let an agent agree plans or release refunds before the people handling its held cases have been assessed on real decisions. A utility can adopt the practice of tying approval rights for each depth to a current certificate.
Utilities that are essential entities under NIS2 in the EU also carry cybersecurity risk management and incident reporting duties, so reviewers should know how to report an agent behaving unexpectedly, not just how to fix the case in front of them.
What should a utility do now?
Plan reviewer capacity for the surge, not the average. Train a core of reviewers in billing and collections before the agent's depth widens, and pair temporary or seconded staff with them during peaks. Run calibration on past exception cases so different teams apply the same standard. Feed every recurring correction back to the agent's owner.
The order for any organisation is in the guide to training teams for AI agents. For a utility, the AI agent training plan template for utilities turns it into weeks, names and cases, and the companion piece on how utilities should train staff before deploying AI agents covers the first months.
Where the Academy fits
Pryme Intelligence Academy's free readiness assessment suits a utility's billing or collections team of up to 25 people: the team takes the Foundations test set in its industry, and the report shows scores by concept, the gaps, and the first agent to start with. Reviewers then move to the Operator track, which certifies on two weeks of real review decisions.
Questions people ask
Do field engineers need reviewer training?
Only if they approve held actions, such as work orders an agent has drafted. Most need the shared vocabulary so they understand why some jobs arrive drafted and some do not. Train those who decide in the queue to reviewer standard.
Should vulnerable customer cases ever be handled by the agent alone?
The agent can read and flag, and it can draft. Decisions that change how a vulnerable customer is treated should stay with a trained person. Set that as one of the utility's own limits so those actions are always held.
How do we size the reviewer team?
Use the volume of held actions the agent produced during its shadow period, then plan for your largest surge. Add cover for leave and illness so no queue depends on one person.