Reconstruct, then judge. AML investigators review an agent's decisions by rebuilding each case from the record: what alert came in, which facts the agent used and where they came from, which bound decided the verdict. Only then do they judge whether the outcome was right, and approve, edit or decline it with a reason a supervisor could follow.
What makes AML review different from other review work?
Two features. First, the costly error is usually the quiet one. An agent that wrongly escalates a case wastes an investigator's time; an agent that wrongly downgrades or closes one lets suspicious activity pass without anyone noticing. Review in AML has to look hardest at what the agent decided not to escalate.
Second, the reasoning is the product. A supervisor examining a case file reads the narrative and the rationale for each step. An investigator who approves an agent's draft adopts its reasoning, so they must be able to defend every sentence.
What are the review steps?
- Open the record and read the verdict. Allowed, held, refused or withheld. For held items, the investigator decides; for allowed items, they test.
- Find the deciding bound. The minimum chain names which of the five bounds decided. A hold on the customer's own limits is a policy question; a hold on the source bound means the agent lacked a fact it needed.
- Rebuild the evidence. List the facts the agent cited and mark each one's provenance. Ledger transactions are sealed. Identity documents verified by a provider are signed. Customer explanations are declared. Adverse media is external.
- Test the reasoning. Does the typology the agent applied fit? Does the narrative explain the activity, or only describe it?
- Look for what is missing. Linked parties, prior alerts, relationship manager notes, other products the customer holds.
- Decide and write the reason. Approve, edit or decline. In AML, the reason should name the evidence that decided it.
- Record any pattern. If the agent repeatedly misreads a typology, that is a change to its brief or sources, routed to its owner.
What may the agent do, and what stays with the investigator?
| Work | Agent | Investigator |
|---|---|---|
| Enrich alerts with account and transaction data | Depth 1, always | Relies on it, checks gaps |
| Prioritise the alert queue | Depth 3 if certified for it | Samples the low-priority end |
| Draft case narratives and closure rationales | Depth 2 | Edits and owns them |
| Close alerts | Held, or depth 3 for defined low-risk types | Decides or samples |
| Decide to report suspicion | Never | Investigator and reporting officer |
| Contact the customer | Draft only, always held | Checks tipping-off risk |
The last two rows are the core of what AML investigators should never let an AI agent do.
An example: testing the bottom of the queue
Take a payments firm where the agent prioritises alerts and, for one narrow alert type, closes them at depth 3 with sampling. An investigator pulls a random sample of 30 closed alerts for the month. Twenty-nine are sound. One closure rested on the agent matching a customer's transfers to a declared business purpose; the ledger shows the counterparties are unrelated to that purpose.
The investigator reopens the alert, records the reason, and escalates it. They then check whether the same reliance on declared purpose appears in other closures. It appears in four more, two of which need reopening. The owner narrows the agent's closure scope, and the business re-certifies it before widening again. The review found a pattern, not just a bad case. That is the argument behind why AML investigators are the real control in an AI agent deployment.
Which controls make review reliable?
- Sampling targets closures and low priority, not only escalations.
- Second review for a share of investigator approvals of agent drafts.
- Reasons cite evidence, not "agree with agent".
- Approval rights follow certification. A business can require a current reviewer certificate, such as Certified Pryme Operator from Pryme Intelligence Academy, before an investigator approves closures.
- Time per case is monitored. A falling average suggests skimming.
Market differences matter. Compare the briefs on AML investigators in payments in Canada and in Brazil. Whether the role survives the arrival of agents is answered in will AI agents replace AML investigators, and the role map in AI agent skills by role.
Questions people ask
How large should the closure sample be?
Large enough to detect a pattern within a month, and larger after any change to the agent, its typologies or its sources. Record the sample method so it can be audited.
Who reviews the investigators' approvals?
A senior investigator or quality team, on a sample basis. The agent's record makes this easier, because the whole case can be reconstructed.
Can an agent check another agent's AML decisions?
It can flag inconsistencies at depth 1. Judging whether suspicion exists stays with people.