Utilities should build agent training around the decisions that can harm a household: disconnection, debt recovery, priority services for vulnerable customers and estimated bills. This template assigns each team a track, a training window that avoids winter peaks and outage periods, and the evidence each person leaves, so the people approving held actions can be shown to be competent.
What is this plan for?
A utility agent does a lot of routine work well. It explains bills, takes meter readings, books field visits, updates contact details and drafts letters about payment plans. Most of that sits at depth 2 or 3.
A small set of actions carry real harm. Restricting supply or starting a disconnection process for non-payment. Removing a customer from the priority services register. Writing off or re-billing a large back-dated charge. These belong at depth 4, held for a named person, or refused entirely by the utility's own limits. The plan exists to make sure the people deciding those actions are trained, assessed and current.
Use it before the agent handles any customer in arrears, and before it touches priority services data. The general sequence for all teams is in the guide to training teams for AI agents.
How do you fill it in?
Start from the agent's capabilities, team by team, and mark the depth of each. Then fill in, for each team:
- the held actions and refusals it will decide;
- which people are on the held rota and so need certifying;
- the window, avoiding the winter billing peak and any period of planned network work;
- what the practicum must include: at minimum, a vulnerable customer case and a payment arrangement outside bounds;
- the evidence that will sit in the register.
Debt teams deserve the most care. A collections agent that offers a payment plan within bounds is useful; one whose reviewer cannot recognise a customer in difficulty is a liability. The training plan template for collections agents goes deeper on that role. For the wider preparation, read how utilities should train staff before deploying AI agents.
The template
| Team | Agent work (depth) | People trained | Track | Window | Practicum must include | Evidence |
|---|---|---|---|---|---|---|
| Billing | Bill explanations (2); meter reading updates (3); re-bills above limit (4, held) | Billing reviewers | Certified reviewer | Spring or summer | A disputed estimated bill | Practicum record; certificate |
| Debt and collections | Payment plan offers within bounds (3); supply restriction steps (4, held) | All approvers | Certified reviewer | Before any arrears contact | A customer disclosing difficulty | Practicum record; certificate |
| Priority services | Reads register (1); drafts updates (2) | Register owners | Certified reviewer | Before any register access | A removal request from a third party | Practicum record; certificate |
| Field scheduling | Books and moves visits (3) | Schedulers | Foundation | Any quiet period | Not required | Test score |
| Outage communications | Drafts customer notices (2) | Comms staff | Foundation | Before storm season | Not required | Test score |
| IT and integration | Connections to billing and metering systems | Systems team | Builder | Before certification | Certified agent | Certified agent on utility cases |
| Agent owner | Depth decisions | Head of customer operations | Architect | Before any depth 4 capability | Not applicable | Signed depth decision |
What does a filled-in plan look like?
Consider an illustrative regional water and energy supplier with a customer operations team of 150. Phase one puts the agent on billing queries and meter readings.
All 150 staff take the shared foundation over three weeks in late spring. The 30 billing reviewers on the held rota take the reviewer track and finish a two-week practicum by midsummer, deciding real held re-bills and refusals. Debt is phase two: the agent gains a payment plan capability only after the 18 debt approvers are certified, and their practicum includes a customer who mentions a medical device at home. That case must route at once to the priority services team, and the reviewer is marked on whether it did.
The systems team of four takes the builder track alongside. The head of customer operations signs the depth decision for each capability, and renewal for everyone falls before the next winter peak. Total cost by phase is set out in how much AI agent training costs for utilities.
Where do utilities usually go wrong?
They treat field and comms staff the same as debt approvers, giving everyone one short course. Or they let the agent's debt capability go live on the promise that reviewers will be trained later. Both leave the hardest decisions with the least prepared people. More on this in what utilities get wrong about AI agent training.
Questions people ask
Should the agent ever start a disconnection process itself?
Not on its own authority. Treat every step towards restricting supply as depth 4 and held for a certified person, and narrow it further through your own limits if your policies require.
How do you train for vulnerable customer cases without real ones?
Use real ones from the held queue during the practicum, with the agent at depth 2 so nothing goes out unchecked. Add past cases, anonymised, to calibration sessions.
Do contractors in field teams need training?
Only if they decide or act on the agent's output. A contractor who receives a booked visit needs nothing new; one who approves a reschedule the agent held needs the foundation at least.