AI agent training for customer support agents in retail banking in Australia

Scam reports, card disputes and hardship requests are where Australian retail bank support staff most need training before AI agents join them. They must know each agent's depth, read held and refused verdicts, check transactions at source, and keep scams, hardship, complaints and fee refunds with people. APRA, ASIC and the OAIC set the expectations.

What does a customer support agent at an Australian bank handle?

In this brief, "support staff" means the people and "the agent" means the AI agent. Support staff in a bank's contact centre, chat team or branch phone line handle balance and transaction questions, lost and stolen cards, disputed card transactions, fast payment problems, fee queries, access and login issues, and changes of personal details.

The hardest contacts are the ones that matter most. A customer who has just sent money to a scammer needs the bank to act within minutes. A customer struggling with repayments may mention it only in passing. A customer describing pressure from a partner over their account may be experiencing financial abuse. Support staff must recognise all of these, often inside an ordinary-sounding query.

Which tasks can the AI agent take, and at what depth?

Certification uses the bank's own past contacts, including scam reports buried inside routine queries, so the depth the agent earns, from 1 (read) to 4 (moves money or is hard to undo), reflects real traffic. The role's ceiling caps it.

What should support staff never let the AI agent do?

For an Australian bank, the general list in what customer support agents should never let an AI agent do translates into five things the agent must never do:

• handle a scam report beyond taking the first details and handing over;
• decide anything about hardship or financial difficulty;
• change a customer's contact details, which fraudsters use to take over accounts;
• tell a customer a complaint has been resolved;
• move money between accounts or reverse a payment.

Some of these are the bank's own limits, which make the action refused for the agent. Others sit at the platform floor. Reversing a settled payment on Pryme's own rail, for example, is withheld for every agent at every depth.

What must support staff learn?

They need to read a verdict: allowed, held, refused or withheld, with the record naming which of five bounds decided, namely role scope, certificate, sources, the bank's limits or the platform guardrails. They need provenance: the core banking ledger is sealed; a merchant's statement in a dispute comes from a named party; a customer's account of what happened is declared; a forwarded text from someone claiming to be the bank is external, and may be the scam itself.

They need to recognise held actions that should not be released, such as a fee refund requested by a caller who failed identity checks. They need to know the agent's lifecycle: its certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day. And they need to know where they sit; an AI agent competency matrix for support leadership shows which capabilities each person is trained to oversee.

Which Australian rules shape oversight and literacy?

APRA supervises banks as authorised deposit-taking institutions. CPS 230 on operational risk management requires banks to manage operational risk, maintain critical operations within tolerance and manage material service providers, which covers the systems behind agent-assisted support. ASIC administers conduct rules, including dispute resolution requirements that set expectations for recognising and responding to complaints. An agent that misses a complaint creates a breach that support staff must catch.

The Privacy Act 1988, overseen by the OAIC, governs personal information, and bank customer data is among the most sensitive. Banks face growing expectations to detect, prevent and respond to scams, and the industry's banking code adds commitments on vulnerable customers and financial difficulty. The government's voluntary AI safety standard sets out practices such as human oversight and transparency. The wider position is in how AI agents are regulated in Australia.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

A 90-day plan

1. Weeks 1 to 4. Literacy for every support staff member: depth, verdicts, provenance and the record, with role plays on a scam report hidden inside a balance query.
2. Weeks 5 to 8. Team leaders review held fee refunds and dispute drafts daily, and calibrate weekly on complaint recognition.
3. Weeks 9 to 12. Release rights tie to a current certificate. Rehearse a certificate lapse at a weekend peak.

Scam reports peak in the evenings and at weekends, so a reviewer rota for support leadership should put certified reviewers on those shifts first. Branch staff and collections teams need training too, as which roles in retail banking need AI agent training explains; the guide to AI agent skills by role has the full map.

Where the Academy fits

Pryme Intelligence Academy sets Foundations tests in retail banking and in the learner's role, so support staff practise on scam, dispute and hardship scenarios. Team leaders take the Operator track: a written exam with 32 of 40 correct to pass, and a practicum of two weeks of real review decisions with no missed red line. A bank can make release rights depend on a current certificate.