Vulnerable customers and essential services set the tone. A utility's programme should train billing, debt, field scheduling and customer teams on the agent decisions they will approve, with heavy practice on customers in vulnerable circumstances, drills for outages and agent failures, a common grounding for leadership, and sign-off that expires and is renewed.
What makes a utility different?
People cannot easily switch off their need for water or energy. A billing error, a missed priority flag or a wrongly scheduled disconnection lands on households who may have no alternative. That raises the stakes for every held action an agent puts in front of a reviewer.
Utilities are also operators of essential services. In the EU, many fall under NIS2 as essential entities, with duties around cybersecurity risk management and incident reporting. Reviewers do not need to know the directive; they need to know what to do when an agent stops, misbehaves or is suspected of being manipulated. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Which teams need which track?
- Billing operations: the reviewer track, with practice on meter read exceptions, back-bills and bill explanations.
- Debt and collections: the reviewer track, with the most practice of any team on vulnerable customers and payment plans.
- Priority services and vulnerability specialists: the reviewer track, because they decide how flags change what the agent may do.
- Field scheduling and dispatch: the reviewer track, focused on held work orders and safety jobs.
- Customer service team leads: the reviewer track for those who approve agent replies at volume.
- Leadership, risk and the security team: the common grounding, with the security team also learning how an agent's actions appear in the record.
- Whoever builds and certifies agents: the builder track.
What may the agent do, and what stays with staff?
| Work | The agent | Depth | Stays with a person |
|---|---|---|---|
| Meter reads | Validates reads, flags exceptions | 1 to 2 | Accepting a read that changes a bill materially |
| Bills | Drafts explanations and corrected bills | 2 | Any back-bill, and any bill to a flagged customer |
| Payment plans | Proposes plans within set rules | 3 | Plans for customers in difficulty |
| Field work | Books routine visits | 3 | Safety jobs and anything involving a supply interruption |
| Outage notices | Drafts updates from network data | 2 | Publication |
| Disconnection | None | Not delegated | Every step |
An example: a catch-up bill to a flagged customer
Take a regional energy supplier. A customer has sent their own meter readings for a year; a smart meter then comes online. The customer reads are declared provenance; the smart meter data is sealed. The agent finds the two far apart and drafts a catch-up bill for several hundred. The account carries a priority services flag for a medical condition, so the draft is held.
A trained billing reviewer sees the flag, checks whether the old reads might be right and the meter wrong, and routes the case to a vulnerability specialist before anything is sent. An untrained reviewer sees a correctly calculated bill and approves it. Practice built from the utility's own exceptions makes this habit routine.
How should drills, sign-off and renewal work?
Run two drills before go-live: an outage, in which the agent drafts many notices fast and reviewers must keep quality under pressure; and an agent failure, in which a run that clears no depth drops the agent to depth 1 the same day and billing must continue by hand. Agent certificates lapse after 90 days, so this will happen.
Sign-off should be against a fixed standard with no missed red line, checked by someone outside the team, and lapse yearly. A utility can make approval rights depend on a current certificate, depth by depth.
The AI agent training plan template for utilities sets out a schedule, and what utilities get wrong about AI agent training covers the traps. Large, shift-based teams can scale with a train-the-trainer programme. Utilities with a retail finance arm will find parts of the programme for banks useful. The general method is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy sets module tests in the learner's own industry and role, so a debt team lead is tested on utility debt cases. The Operator certificate requires a two-week practicum with no missed red line, signed off by a Pryme assessor, and renewal after 12 months is an update module and a short quiz.
Questions people ask
Should field engineers be trained?
Engineers who receive agent-booked jobs need a short briefing on what the agent may change. Those who approve work orders need the reviewer track.
How do vulnerability flags affect agent depth?
They should lower it. A business can set its own limits so that any action on a flagged account is held for a person, whatever depth the agent has earned elsewhere.
What should the security team learn?
How agent actions appear in the record, how to spot unusual patterns, and that some actions are withheld at the platform floor whatever the settings.