Brazilian payment institutions should train AML investigators to verify the evidence an agent gathers on Pix alerts, read the network links it draws between front accounts, review its draft reports and return requests, and decide every block themselves. The suspicion and the report remain the investigator's; the agent prepares the file faster and records how.
What fills a payments AML investigator's day in Brazil?
Pix changed the rhythm. Money moves instantly, at any hour, between accounts that can be opened in minutes. Investigators at payment institutions spend much of the day on alerts about accounts receiving many Pix transfers from unrelated people and passing them on at once, accounts opened with borrowed or bought identities, and the front accounts, the contas laranja, that launder scam proceeds.
Alongside alert work, investigators handle fraud-return requests under Pix's special return mechanism, decide on precautionary blocks of received funds, respond to court and police orders, and prepare communications to Brazil's financial intelligence unit. Each decision is time-pressured: a block that comes an hour late may find the account empty.
Which parts of the work can an agent take, and how far?
Alert enrichment (depth 1). The agent assembles the customer's registration data, account age, Pix key history, device and IP data, inflows, outflows and previous alerts. Each fact carries its source, so the investigator can tell what the payment rail recorded from what the customer declared.
Front-account network analysis (depth 1). The agent traces common devices, keys, funders and beneficiaries across accounts. Networks are where front accounts reveal themselves, and mapping them by hand is slow.
Return request and block drafting (depth 2 draft, depth 4 decision held). The agent drafts the return request or the precautionary block with the evidence attached. Blocking or returning funds affects a customer and is hard to undo, so a named investigator decides.
Report drafting (depth 2). Where the investigator has decided to report, the agent drafts the communication with the facts in a consistent structure. The investigator edits and submits.
Case closure notes (depth 2). The agent drafts the reasons for closing an alert. The investigator signs off.
What must investigators never let the agent do?
The agent must never tell a customer, directly or by implication, that their activity is under review or has been reported; Brazilian AML law forbids disclosing that a communication has been made. It must never release a block or reverse a return on its own, and never treat a customer's explanation as fact without checking it against sealed transaction data. A declared explanation that the agent accepted unverified is the most common weakness in an agent-drafted case. The complete list, with reasons, is in what AML investigators should never let an AI agent do.
These are customer-set limits: the institution narrows what its agent may do, and those actions are refused or held. Separately, the platform itself withholds certain actions from every agent at every depth, such as reversing a settled payment on its own rail.
What does the investigator need to learn?
Three skills matter most. Reading a verdict: which of the five bounds decided (role scope, certificate, sources, the institution's limits, the platform guardrails) and what that means for the case. Reading provenance: sealed facts from the ledger and the rail outrank anything the customer declares, and external material such as a social media post can prompt a question but cannot decide one. And correcting well: when the investigator overrules the agent on a pattern, such as informal community savings groups that look like structuring, the correction should change the agent's brief. Banking corrections that improve the agent shows how.
The investigator remains the real control in an agent deployment. The work shifts from gathering to judging, which is the theme of will AI agents replace AML investigators?.
Which Brazilian rules shape oversight and competence?
Banco Central do Brasil authorises and supervises payment institutions and runs Pix. Its rules require institutions to keep AML/CFT policies, monitor transactions, assess risk and train staff, and it sets the Pix mechanisms for fraud returns and precautionary blocks. An institution that uses agents in investigations should be able to show the Banco Central who decided each block and on what evidence.
Brazilian AML law requires suspicious operations to be communicated to the financial intelligence unit and prohibits telling the customer. The LGPD governs the personal data the agent processes, and the ANPD enforces it. The LGPD gives people rights in relation to decisions made solely by automated processing that affect their interests, which is another reason to keep blocks and exits with a person.
The FATF Recommendations set the international standard for the risk-based approach and suspicious transaction reporting that Brazil's rules reflect.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
A 90-day plan for the AML team
| Period | Investigators | Agent |
|---|---|---|
| Days 1 to 30 | Foundation learning using the institution's own closed Pix cases | Depth 1: enrichment and network maps; investigators compare with their own work |
| Days 31 to 60 | Review all drafts; weekly calibration on disputed cases; corrections banked | Depth 2: case notes, return requests, report drafts |
| Days 61 to 90 | Investigators who decide blocks complete certification and supervised practice | Blocks and returns stay held; any narrow closure class tested in Attest first |
A ready structure is in the AI agent training plan template for AML investigators. For analysts, compliance officers and fraud teams, see AI agent skills by role.
Where the Academy fits
Pryme Intelligence Academy's Operator track covers how a governed agent decides, human review, teaching the agent and handling incidents, with module tests set in the learner's own industry and role. Certification needs a 40-question exam and a two-week practicum with no missed red line. Certificates last 12 months and can be checked at source through VerifyDoc.ai.
Questions people ask
Can the agent run at night when investigators are off shift?
It can enrich and draft overnight. Blocks and returns should still go to a named on-call investigator, because they are the decisions that need a person.
How should we treat an alert where the customer's story checks out?
Record which sealed facts confirmed the story. The closure note should cite them, so a reviewer can see the explanation was verified, not just accepted.
Should the agent see court orders?
Read-only access to the orders relevant to a case is useful. Responding to an order stays with a person.