Under the FATF risk-based approach, firms must spend their AML effort where the risk is highest, and that includes training. AML staff who work with AI agents need training scaled to the risk of the decisions they make: the analyst releasing an agent-drafted suspicious activity report needs far deeper competence than the colleague who reads an agent's screening summary.
What does the risk-based approach ask of AML staff?
The FATF Recommendations are the international standard for anti-money laundering and counter-terrorist financing. National laws and supervisors put them into effect. At their centre is the risk-based approach: identify and assess your money laundering and terrorist financing risks, then apply measures in proportion, with enhanced measures where risk is higher and simplified ones where it is demonstrably lower.
The standard also expects firms to run AML programmes with internal policies, procedures and controls, including ongoing training for employees. Read together, these two ideas set the training rule: people should be trained in proportion to the risk of what they do. Our explainer on what the risk-based approach is covers the concept itself.
Agents change where risk sits. An analyst who used to gather screening results now reviews an agent's assembled case. The judgement remains, but it arrives pre-packaged, and the analyst must be able to tell when the package is wrong.
Who does it apply to?
The FATF standard reaches financial institutions and certain non-financial businesses and professions, such as lawyers, accountants, estate agents and dealers in high-value goods, as national law implements it. Inside a firm, staff working with AML agents typically include:
- Alert analysts who review transaction monitoring alerts an agent has triaged.
- Customer due diligence teams who release onboarding and periodic review files the agent assembled.
- Enhanced due diligence specialists handling higher-risk customers, politically exposed persons and complex structures.
- Investigators who take escalated cases and decide whether suspicion is formed.
- The money laundering reporting officer or equivalent, who decides whether a report is filed.
- Front-line staff who receive agent prompts to ask a customer for source of funds.
The wider question of how FATF standards apply to AI agents in AML is covered separately. Here the question is what these people must know.
What competence does a risk-based training plan imply?
Map training to the risk of the decision, not the job title. A useful split:
Low-risk exposure. Front-line staff who see an agent's prompt to request documents need to know the agent's role, what a held request means, and the tipping-off rule: they must never tell a customer that a report is being considered.
Medium-risk exposure. Alert analysts and due diligence teams release agent work. They need to understand the agent's certified depth and the depth each action needs: the agent reads at depth 1, drafts a case summary or a request for information at depth 2, and may close clearly false-positive alerts at depth 3 only if it is certified to that depth. They must read provenance: a sealed transaction record outweighs a customer's declared explanation, and an emailed invoice is external and proves nothing alone.
High-risk exposure. Investigators and enhanced due diligence specialists must judge cases the agent cannot settle: layered structures, inconsistent stories, sanctions near-matches. They need practice overturning confident agent conclusions, and they need to know what the agent never saw.
Decision owners. The reporting officer decides whether to file. An agent can draft a report at depth 2. Filing is held for a named person, every time, and that person must be able to read the full record of how the case was built.
Two agent-specific risks deserve their own training. First, an agent closing alerts at depth 3 can repeat one flawed rule across thousands of alerts, so analysts must sample closed alerts, not only open ones. Second, an agent that writes to customers must be bounded so that it never discloses suspicion; staff should know that such a message would be refused and why.
Competence is shown by work, not attendance. Work-based assessment, where reviewers are scored on real or seeded cases, is the evidence that fits a risk-based standard.
How do you meet it in practice?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Risk assessment drives training | Training tiers mapped to the risk of each role's AML decisions | Risk assessment section linking roles to training tiers |
| Ongoing employee training | Role-specific modules on each AML agent, refreshed on change | Training records by person and agent |
| Enhanced measures for higher risk | Deeper training and assessment for EDD and investigations staff | Assessment results; seeded case scores |
| Decisions stay with competent people | Only current, trained staff release held cases and reports | Release log with approver and training status |
| Tipping-off prohibition | Agent bounds refuse disclosure; staff trained on the rule | Refusal records; module completion |
| Effectiveness | Sampling of agent-closed alerts by trained analysts | Sample results and findings |
| Currency | Retraining when typologies or agent bounds change | Change-linked training records |
What will a supervisor or auditor ask?
- How did your risk assessment decide who gets which training?
- Ask an analyst: what does the agent close without you, and how would you know it was wrong?
- Who sampled agent-closed alerts last month, and what did they find?
- Show us the record for this report: who drafted it, who decided to file.
- How do you stop an agent tipping off a customer?
- What changed in training after the last new typology?
What does the standard not cover?
The FATF Recommendations do not bind firms directly; national law and supervisors do. They do not prescribe training hours, formats or certificates. They say nothing specific about AI agents, so firms must apply the principles themselves. And they do not replace data protection law, which governs how customer data reaches the agent. The FATF risk-based approach applied to AI agents covers the system side. Markets differ in how they supervise; the UK's principles-based approach to AI is one example. For structured, checkable training of reviewers, Pryme Intelligence Academy's Operator track uses a practicum of real review decisions. The guide to AI literacy and certification explains the wider model.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Can an agent close false-positive alerts on its own?
Only if it has been certified on your own cases to do so, at depth 3, within bounds you set. Trained analysts should still sample what it closes.
Who decides whether a suspicious activity report is filed?
A named person, usually the reporting officer or a delegate. The agent can draft; filing is held for that person every time.
Does the risk-based approach mean low-risk staff need no training?
No. It means their training can be lighter. Everyone who handles AML-relevant work needs to know their role, the tipping-off rule and when to escalate.