PrymeIntelligence

Training teams for AI agents·Article

What local councils get wrong about AI agent training

What local councils get wrong about AI agent training: one-off sessions on thin budgets, digital teams trained before service staff, members left out.

Councils rarely lack goodwill towards training; they lack time and budget, so they run one session for the digital team and hope it spreads. It does not. Revenues, housing and customer service staff end up approving refunds, repair orders and arrears letters they were never taught to judge, and members cannot get a straight answer about who checks the agent.

Why does training the digital team first go wrong?

The digital or transformation team usually leads the procurement and the pilot, so it gets the most training. It understands the agent's connections, its brief and its settings.

But the council's agents do their work in services. A revenues agent drafts council tax reminders and prepares refunds of overpaid accounts. A housing repairs agent triages reported faults and raises routine work orders within the council's own limits at depth 3, holding anything costly or unusual. A customer contact agent drafts replies on bin collections, parking permits and planning enquiries for staff to send at depth 2. The people who decide the held cases are revenues officers, repairs coordinators and customer service advisers. Train them first, on their own service's cases, and train the digital team as builders and stewards.

What is wrong with a single big session?

It fits the budget and the calendar, and it leaves no trace a month later. Council staff are often part-time, spread across sites, or shared with neighbouring councils through joint services. A half-day session reaches some of them once.

Short, repeated, on-the-job training works better and costs less. Twenty minutes on reading a held refund, followed by supervised decisions on the live queue, teaches more than a morning of slides. Agent certificates lapse after 90 days, and a run that clears no depth drops an agent to depth 1 the same day, so the team will meet surges of held work at unpredictable moments. Staff who have been taught in small, regular pieces handle those surges; staff who sat one session do not remember what to do. A practical plan, with who trains when, is in the AI agent training plan template for local councils.

Which council decisions need the most care?

Three stand out, and generic public sector training rarely covers them.

Refunds to changed bank details. A council tax refund moves money and sits at depth 4, held for a named officer. Requests to send refunds to a new account are a known route for fraud. A bank detail change that arrived by email is external and cannot authorise a payment; the officer must verify through a known channel before releasing anything.

Arrears and enforcement. An agent can draft reminder letters at depth 2. Anything that moves a resident towards enforcement should stay with a person who has checked for vulnerability, disputes and payment arrangements already in place. Officers must be trained to look for signs the agent flagged or missed.

Repairs that affect safety. A repairs agent may classify a reported fault as routine when the tenant's description suggests otherwise. Coordinators must know when to override the agent's priority and how to record why.

Why do elected members need to be part of it?

Members scrutinise services and answer to residents. When something goes wrong with an agent, they will ask who checked it, what the agent was allowed to do and whether the officer was trained. Councils that leave members out of training get those questions at a public meeting and struggle to answer them.

Members need a short session on the vocabulary: depth, the four verdicts, what held and withheld mean, and what the record shows. With that, they can ask sharp questions and recognise good answers. Public bodies also face additional transparency and fairness duties, and in the EU some public uses of AI are treated as high-risk under the AI Act. This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

How should shared services be handled?

Many councils share revenues, benefits or contact centre services with neighbours. If an agent serves more than one council, each council's own limits apply to its residents, and the staff reviewing held actions must know which council's rules govern each case.

Train shared-service staff on the limits of every council they serve, and assess them on cases from each. Keep one register of who is certified, with expiry dates, that every partner council can see. A council can adopt the practice of granting approval rights only to staff with a current certificate, depth by depth, and should expect its partners to do the same.

What should a council do instead?

  1. Train service staff before the digital team, on their own service's cases.
  2. Replace the single session with short, regular training on the live queue.
  3. Drill the three high-care decisions: refunds, arrears and safety repairs.
  4. Give elected members a short session so scrutiny has substance.
  5. Align training and approval rules across shared services.

The broader order is in the guide to training teams for AI agents. For the detail, see how local councils should train staff before deploying AI agents and how much AI agent training costs for local councils. Banks share the refund fraud problem at a larger scale; what banks get wrong about AI agent training covers their version.

Questions people ask

Can a council run agents with only part-time staff as reviewers?

Yes, if the rota covers every hour the agent produces held actions and every reviewer is trained and assessed. Part-time reviewers need the same standard as full-time ones, and more attention to handover between shifts.

Should residents be told when an agent drafted their letter?

Be open about how the council uses agents and who checks their work. Where residents interact directly with an AI system, tell them so. A named officer approving a decision is the reassurance most residents want.

What is the cheapest useful first step?

Train the revenues team on held refunds and bank detail changes. It is a small group, the risk is concrete, and the cases are easy to find in past records.

Train the people who run your agents.

Pryme Intelligence Academy: free courses on the real product, audio lessons, labs in a free workspace, and certificates anyone can check on VerifyDoc.ai. Teams get seats, an admin view and a free readiness assessment.