FATF risk-based approach applied to AI agents

The FATF risk-based approach asks firms to match their effort to their money laundering risk. Applied to AI agents, it gives a second rule: match the agent's authority to the risk of the task. Agents may work freely on low-risk preparation, prepare drafts on higher-risk files, and never take the decisions where an error would let illicit funds move.

What does the risk-based approach require?

The risk-based approach sits at the start of the FATF Recommendations, and it shapes everything after. In substance, firms must:

• identify and assess the money laundering and terrorist financing risks they face, across customers, products, channels and geographies;
• document that assessment and keep it current;
• apply measures in proportion: enhanced where risk is higher, simplified where lower and permitted;
• have policies, controls and procedures approved by senior management to manage the risks identified;
• monitor whether those controls work.

It is a discipline of allocation. Resources go where the risk is. Our explainer on the risk-based approach covers the general idea, and our explainer on customer risk ratings covers how it reaches individual customers.

Who does it apply to?

Every obliged entity under national AML laws based on FATF standards, and supervisors too, who are expected to supervise on a risk basis. That includes banks, payment and e-money firms, crypto-asset service providers, insurers for relevant products and non-financial businesses such as lawyers, accountants and dealers in high-value goods.

What does it mean for where agents are used?

The risk-based approach usually answers "how much due diligence?" With agents, it answers a second question: "how much may the agent do?". The same logic applies. Low-risk work can be delegated further; high-risk work needs more human attention.

That cuts against a common instinct. Firms often want agents on their highest-volume, lowest-value work, then gradually extend them to complex cases as confidence grows. The risk-based approach says the complex cases are where human attention must concentrate, so agents there should prepare, not decide. The time agents free up on routine work should be spent by people on the high-risk files.

Here is how authority can follow risk across common AML tasks:

• Collecting documents from a low-risk retail customer. An error is low-risk and caught later in review. Depth 3: the agent requests, chases and files within set rules.
• Extracting data and screening a standard business customer. A missed hit is possible. Depth 2: the agent prepares results for an analyst.
• Summarising a transaction alert. Framing can bias the analyst. Depth 2: the agent cites facts and recommends; the analyst decides.
• Mapping ownership of a complex structure. Hidden owners are the whole point of such structures. Depth 1 or 2: the agent assembles sources; the analyst concludes.
• Spotting trade-based laundering patterns. This needs context and judgement. Depth 2: the agent flags anomalies with evidence.
• Deciding to report or exit a customer. The highest risk. No authority: the decision is held for a named person.

Our short answer on whether an agent can spot trade-based money laundering explains why pattern work in trade finance needs a person to interpret.

How do you apply it to the agents themselves?

Treat each agent as something your risk assessment must cover, and set its bounds from that assessment.

What will a supervisor ask?

• How does your risk assessment account for the agents you use?
• Why does this agent have authority to act on this task but not that one?
• Who approved the agents' bounds?
• Are your most experienced analysts spending more time on high-risk work since agents arrived?
• How do you know the agents perform as well on higher-risk tiers as on lower ones?

What does the risk-based approach not do?

It does not let a firm lower its standards because an agent makes them cheaper to meet. It does not remove fixed obligations, such as reporting suspicion or applying sanctions without delay, which apply whatever the risk score. And it does not govern the AI itself; model risk and AI rules do that. The UK's regulator-led approach to AI follows a similar philosophy, which we discuss in the UK principles-based approach to AI. For how agents support monitoring once bounds are set, see FATF ongoing monitoring and AI agents, and for the principle behind it all, our guide to AI compliance agents.

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

Where Pryme fits

Pryme sets the depth each action needs, so the same agent can chase documents at depth 3, draft alert summaries at depth 2 and be refused any reporting decision. The agent's role sets a ceiling, Attest evidence sets the depth, and the lower wins. That gives supervisors a record of authority that follows risk, task by task.