PrymeIntelligence

AI agent skills by role·Sector brief

AI agent training for AML investigators in payments in South Africa

How South African payment firms should train AML investigators to oversee AI agents on mule, SIM-swap and cross-border alerts, with a 90-day plan.

South African payment providers should train AML investigators to verify the evidence an agent assembles, recognise when it misreads legitimate community patterns such as stokvels, review its draft reports, and decide every restriction and report themselves. Calibration across the team keeps the agent learning one standard rather than several. The suspicion decision never moves to the agent.

What does AML investigation involve at a South African payments firm?

Third-party payment providers, wallet operators, cross-border remittance firms and payment facilitators see a wide range of activity. Investigators look for mule accounts receiving proceeds of SIM-swap and phishing fraud, merchants whose turnover does not match their trade, remittances to neighbouring countries that do not fit the sender's profile, and cash-heavy activity that may be informal trade or something worse.

Legitimate patterns often look suspicious. A stokvel, a community savings club, collects regular contributions from many members into one account and pays out in a lump. Burial societies and church collections behave similarly. An investigator who does not recognise them will file weak reports and alienate good customers; one who over-trusts them will miss a mule network dressed up as a club. Investigators also respond to requests from the financial intelligence unit and law enforcement, and prepare reports.

Where do agents help, and how far can they act?

Alert enrichment (depth 1). The agent assembles identity verification, account and wallet history, devices, SIM-change signals where the firm has them, counterparties and prior alerts. Each fact carries its provenance.

Community scheme recognition (depth 2). The agent compares an account's pattern with the profile of a stokvel or burial society and drafts an assessment, noting which facts support it. A constitution or member list the customer provided is declared, not verified, and the agent should say so.

Mule network mapping (depth 1). The agent links accounts that share devices, funders or rapid onward transfers, especially after fraud reports.

Report drafting (depth 2). After the investigator decides to report, the agent drafts the narrative and lists the facts and their sources. The investigator edits and submits.

Restriction proposals (depth 4, held). Restricting an account stops wages, grants or club payouts reaching people. The agent drafts the proposal; a named investigator decides.

What should investigators learn before working with agents?

Begin with what AML investigators should learn before working with AI agents: depth, verdicts and the five bounds every action meets, from role scope to the platform guardrails. Then provenance, which decides most cases here. A stokvel's member list is declared; regular deposits from many named members over months are sealed ledger facts. The second kind carries far more weight.

Then calibration. If one investigator treats every collective account as a stokvel and another treats every one as a possible mule, the agent will learn contradictory rules from their corrections. Regular calibration sessions for AML investigators agree the standard before corrections are banked. Banking them is the last step: teaching AML investigators to bank corrections explains how a correction becomes a change to the agent's brief or sources.

What stays with the investigator?

Suspicion, the decision to report, restricting or exiting a customer, and every conversation with a customer under review. South African law prohibits disclosing that a report has been made, so the agent's customer-facing drafts are held for a person. Responding to law enforcement and the financial intelligence unit also stays with people.

These are judgements that need local knowledge and accountability, which is why AML investigators are the real control. The work moves from gathering to judging, the shift discussed in will AI agents replace AML investigators?.

Which South African rules shape oversight and training?

South Africa's financial intelligence law requires accountable institutions to maintain a documented risk management and compliance programme, carry out customer due diligence, report suspicious and unusual transactions to the national financial intelligence unit, and train staff to comply. When agents take part in investigations, the programme should describe how people oversee them and the training records should show that they were trained to.

The South African Reserve Bank and its Prudential Authority supervise banks, and the payment system falls under the Reserve Bank's oversight; payment providers often operate through or alongside a sponsoring bank, which will expect sound AML controls. The FSCA regulates market conduct of financial institutions. The Protection of Personal Information Act (POPIA), enforced by the Information Regulator, governs the personal information the agent processes, and restricts decisions about people based solely on automated processing, another reason restrictions are decided by a person.

The FATF Recommendations set the international standard on risk-based supervision and reporting that South African law reflects.

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

A 90-day plan

Period Investigators Agent
Days 1 to 30 Foundation learning with the firm's own closed cases, including community schemes Depth 1: enrichment and network mapping
Days 31 to 60 Review every draft; weekly calibration on collective accounts and mule cases Depth 2: scheme assessments and report drafts
Days 61 to 90 Investigators who decide held items complete certification and supervised practice; programme document updated Restrictions stay held; any new depth tested in Attest first

For how analysts, compliance officers and fraud teams work around investigators, see AI agent skills by role.

Where the Academy fits

Pryme Intelligence Academy's free readiness assessment, for teams of up to 25, has the team sit the Foundations test set in its industry and reports scores by concept, the gaps and the first agent to start with. Investigators who approve held items then take the Operator track and practicum. Certificates are issued through VerifyDoc.ai and checkable at source.

Questions people ask

Should the agent treat all stokvels as low risk?

No. It should assess each against the evidence and say what it relied on. A well-documented club with long, regular history is different from a new account with a club name.

Can the agent use SIM-swap signals?

Where the firm receives them lawfully, they are useful enrichment. The agent should show them as evidence, not treat them as proof of fraud.

Who signs off the updated compliance programme?

The firm's compliance function and senior management, as with any change to the programme. The agent's role should be described in plain terms.

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.