The case for certifying AML investigators before AI agents go live

Certify AML investigators before agents go live because the decision that matters most in their work, whether to report suspicion, never moves to the agent. What moves is everything around it: alert triage, evidence gathering, the draft narrative. An investigator who trusts that preparation without knowing how it was built will file, or fail to file, on someone else's reasoning.

What does an agent take off an AML investigator's desk?

An AML investigator spends most of the day assembling, not deciding. An alert fires. They pull the customer profile, the transaction history, the counterparties, any previous alerts, any adverse media. They write it up. Only then do they make the call: close with rationale, escalate, or recommend a suspicious activity report.

A governed agent can do much of the assembly. At depth 1 it reads the case management system, the core ledger and the screening results. At depth 2 it drafts the case summary and a proposed disposition with the evidence attached. In some deployments it may close a narrow class of obviously false alerts at depth 3, within rules the compliance function sets and Attest has proven on the firm's own past cases.

What it does not do is decide that something is suspicious, or decide that it is not, for anything outside that narrow class. Those dispositions are held for a named investigator. The filing decision itself sits with the person and the firm's nominated officer.

Why is a held disposition harder to review than a blank case?

An investigator starting from a blank case builds their own picture. An investigator reviewing an agent's draft starts from someone else's picture, and the risk is that they check whether the draft is coherent rather than whether it is complete.

That risk is specific to AML. A draft narrative can be well written, accurate in every sentence and still miss the one counterparty that changes the case. Under the FATF risk-based approach, the quality of an investigation depends on whether the investigator asked the right questions about this customer's risk, not on whether the write-up reads well. The investigator has to know what the agent looked at, what it could not see, and which facts were sealed by the ledger versus declared by the customer at onboarding.

Certification tests exactly that. Can the investigator see which sources the agent drew on? Can they spot that a key fact is declared rather than sealed? Can they tell when the agent's proposed closure rests on a missing check? A person who passes those tests can review a draft. A person who has not been tested will, under queue pressure, start agreeing with it.

What does certification need to prove for this role?

For AML investigators the bar is high because the red lines are serious. A useful certificate shows four things.

1. They read the decision chain. For any held case, they can say which bound decided the hold: the agent's scope, its certificate, its sources, the firm's limits or the platform floor.
2. They know what the agent may never do. It may not decide to file or not file outside the narrow class it is certified for. It may not contact the customer about the alert, because that risks tipping off.
3. They catch what is missing. In a practicum of real review decisions, they spot absent evidence, not just wrong evidence.
4. They write their own rationale. Their closure reasoning is their own, recorded with the verdict, not a copy of the agent's draft.

A practicum log is the natural evidence for points three and four; the practicum log template for AML investigators shows what to capture.

Doesn't the investigator's existing expertise cover this?

Partly. An experienced investigator already knows typologies, red flags and what a good file looks like. Certification does not re-teach that. It teaches the part that is new: how a governed agent reached its draft, where its authority ends, and how to read the record it leaves.

The expertise gap actually runs the other way for some seniors. Experienced investigators can be the most likely to skim an agent's draft, because it looks like what they would have written. Testing them is not an insult; it checks that their judgement is being applied to the case, not to the prose.

There is also the question of standing over time. The agent's own certificate lapses after 90 days and must be renewed against fresh cases, and its depth can drop the same day a run fails. Investigators need to understand what that means for the queue, and the piece on preparing AML investigators for the 90-day recertification cycle covers how to keep the team in step.

What is the trade-off?

Certifying the whole investigations team before go-live is slow, and AML teams are usually short of people already. Alert backlogs are the reason firms buy agents in the first place, and asking investigators to sit exams and a practicum while the backlog grows feels backwards.

A workable compromise is staged. Let the agent run at depth 1 and 2 from the start, gathering evidence and drafting, because that already saves time. Hold every disposition for a person. Certify investigators in waves, and only allow the agent's narrow depth 3 closures in a team where the reviewers who sample those closures are certified.

The reason to accept the delay is accountability. The record will show which investigator closed each case. If a closure later proves wrong, the questions in whether AML investigators are accountable for an agent's mistakes become very practical, and a firm that can show its investigators held current standing is in a far stronger position.

What should a head of financial crime do now?

Write certification into the deployment plan for investigations, with named dates. Decide which dispositions the agent may draft and which it may never propose. Make a current certificate a condition of signing off agent-prepared closures. Keep the practicum logs and exam results where an examiner can see them. And compare notes with colleagues running the same exercise in neighbouring functions; the reasoning for collections agents and for financial controllers rests on the same principle of tested judgement before authority. The guide on AI literacy and certification for people who work with agents lays out the levels.