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

Pix has made Brazilian banking instant, and with it the scams. Support staff at retail banks need training to supervise AI agents that answer account, card and Pix questions: knowing each agent's depth, reading verdicts, checking records at source, and keeping fraud reports, returns of Pix funds and complaints with people, under the central bank's rules and LGPD.

What fills a support agent's shift at a Brazilian bank?

In this brief, "support staff" are the people and "the agent" is the AI agent. Support staff at a Brazilian bank's contact centre and digital channels answer questions about Pix transfers and keys, limits, card bills and disputes, account access, payroll-deductible loans, fees and digital account opening. Much of the contact comes through messaging apps and in-app chat.

Fraud dominates the difficult contacts. A customer may have sent a Pix to someone impersonating a relative on a messaging app, or to a fake seller, or been pressured through a fake call from "the bank". Retirees often call about payroll-deductible loans they did not request. Support staff must act quickly, record the facts precisely and start the right process, while the customer is often distressed.

Where does the AI agent help, and how deep can it go?

Pix scam contacts belong in the certification set from the start. The depth each task earns, from 1 (read) to 4 (moves money or is hard to undo), comes from those past cases, and the role sets the ceiling.

1. Balance, statement and Pix history queries, depth 3 once certified. Answers come from the core ledger.
2. Pix limit and key management guidance, depth 2. The agent explains and drafts; changes to limits follow the bank's authenticated in-app flow, not the conversation.
3. Card bill and dispute intake, depth 2. The agent drafts the dispute from the transaction record and the customer's account.
4. Fraud and scam reports, depth 1. The agent takes first details, flags urgently and hands over.
5. Fee refunds, depth 4, held. A named team leader releases above the bank's limit.

Which skills matter most for support staff?

Which AI agent skills matter most for customer support agents ranks the skills in general. At a Brazilian bank, five stand out.

• Reading a verdict. Allowed, held, refused or withheld, and which of five bounds decided: role scope, certificate, sources, the bank's own limits or the platform guardrails.
• Provenance. The core ledger is sealed. A customer's description of a scam is declared. A screenshot of a messaging app conversation is external evidence, useful but never enough to authorise anything.
• Speed with care on fraud. Knowing exactly what to capture and who to hand to, so a request to return Pix funds starts without delay.
• Banking corrections. Fixing the agent's sources when the same error repeats.
• The lifecycle. An agent's certificate lapses after 90 days, and a run that clears no depth drops it to depth 1 the same day.

What stays with people?

Write who may release what in an approver authority matrix for support leadership.

Which Brazilian rules shape oversight?

The Banco Central do Brasil supervises banks' conduct, operational risk and customer service, including the requirement for banks to maintain an ombudsman unit for unresolved complaints. It operates Pix and its special return mechanism for cases of fraud, which lets the paying bank request the return of funds from the receiving bank within set periods. Support staff must know how to trigger that process and what to tell the customer, and must not let an agent promise an outcome.

LGPD, enforced by the ANPD, requires a lawful basis and purpose for processing, minimisation and security, and gives customers rights, including the right to request review of decisions taken solely on automated processing that affect their interests. Bank data is sensitive in practice, and staff must not copy it into personal messaging apps. The guide to LGPD and training staff who use AI agents goes further. Consumer protection law requires clear information and fair treatment.

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

A 90-day plan

Weeks 1 to 4. The whole floor learns depth, verdicts, provenance, the record and LGPD handling, then rehearses the common Pix scams until staff recognise them inside ordinary queries.

Weeks 5 to 8. Team leaders review held fee refunds and dispute drafts daily. Weekly calibration on fraud handovers checks that nothing was missed or delayed.

Weeks 9 to 12. Release rights tie to a current certificate. Rehearse a certificate lapse on a weekend evening, when scam reports peak. Track progress in an AI agent competency matrix for the support team. Pryme Intelligence Academy's Operator track is one way to certify team leaders.

Payroll-deductible loan specialists and ombudsman staff also handle cases the agent flags, and which roles in retail banking need AI agent training covers them. The guide to AI agent skills by role sets out the rest.