Why AML investigators are the real control in an AI agent deployment

In financial crime compliance, an alert is not a control; a decision is. AI agents can triage alerts and draft the case, but the control a supervisor relies on is the investigator's disposition: suspicious and reported, or not, with reasons. That is why AML investigators remain the real control, and why their training has to change.

What does the agent change, and what does it leave alone?

The investigator's raw material used to be a long queue of alerts, most of them false, each needing the same evidence gathered before any thinking could begin. A governed agent takes that gathering. It reads the alert, the customer's profile, the transaction history and the counterparties. It drafts a narrative and a proposed disposition with its sources attached, each stamped with its provenance. In some deployments it closes a narrow class of clearly false alerts at depth 3, within rules set by compliance and proven on the firm's own cases.

What the agent leaves alone is the judgement at the centre: is this activity suspicious? Under the FATF Recommendations, firms must report suspicious transactions, and that obligation rests on people forming a view. An agent can propose; it cannot form the view the law expects.

So the investigator's queue shrinks to the cases that need a mind. Every one of them is a decision the firm will have to stand behind.

Why is the investigator the deciding control?

Because everything before the investigator narrows the field and everything after depends on their call.

The monitoring rules decide what becomes an alert. The agent decides which alerts are clearly false and what evidence to gather for the rest. Those are both controls, but they are filters. The investigator is the first point where someone decides what the activity means. If they close a case that should have been reported, nothing downstream will catch it; the nominated reporting process never sees it.

The investigator also controls the quality of the filters. When an agent's closures are sampled and an investigator finds one that should have been escalated, that finding narrows the agent's closure class or triggers a review of its certification. When investigators repeatedly find a monitoring rule produces noise, they inform its tuning. The control loop runs through them.

How do investigators review an agent's work well?

By reading the draft as a starting point, not a conclusion. The guide on how AML investigators review an agent's decisions covers the method in detail. In short:

• Read which bound decided the hold, and what it tells you about where the agent's authority ended.
• Check provenance on every fact the disposition rests on. A source of funds that is declared by the customer is not the same as one evidenced by the firm's own records.
• Look for what the agent could not see: other accounts, related parties, information held in another team's system.
• Write your own rationale, recorded with the verdict.
• Sample the agent's depth 3 closures, not only its held cases.

Does speed change the job in some markets?

Yes. In Brazil, Pix instant payments move money in seconds, around the clock. An investigation that takes days may be looking at funds that left long ago, and the pressure to decide quickly is intense. Investigators there must judge agent drafts fast without letting speed become agreement; AI agent training for AML investigators in payments in Brazil covers what that means alongside the Banco Central do Brasil's expectations and the LGPD.

Will agents eventually replace investigators?

Not in the sense that matters. Agents will take more of the gathering and more of the clearly false alerts. The decision on suspicion, and accountability for it, stays with people. The article on whether agents will replace AML investigators makes the argument in full. What changes is that each investigator handles harder cases and has more influence over how the agent behaves.

How does this compare with other roles?

The pattern is shared across functions. Customer support agents become the control on refunds and account changes an agent holds. IT administrators become the control on access an agent proposes. AML is distinctive because the investigator's decision carries a legal reporting obligation and because the people on the other side are actively trying to avoid detection.

What is the trade-off?

Treating investigators as the deciding control means investing in them heavily: deeper training, observed practice, regular sampling of their decisions. That is costly in a function already short of people. It also slows the gains from the agent, since investigators must review its closures as well as their own queue.

The alternative is to treat agent drafts as near-final and investigators as checkers. That is faster and it fails in exactly the cases that matter: the well-written draft that misses the one counterparty that changes everything. Supervisors will look at how dispositions were reached, and a firm whose investigators approved agent drafts without independent judgement will struggle to show its control operated.

What should a head of financial crime do differently?

Name the investigator as the control in your risk assessment, not the agent. Train investigators specifically on reading agent drafts: provenance, gaps, the decision chain. Require observed practice before they sign off agent-prepared dispositions. Build sampling of depth 3 closures into the weekly routine. And make sure investigator findings feed back into the agent's closure rules and the monitoring rules. The guide to AI agent skills by role sets out what each role in the chain needs.