Are AML investigators accountable for an AI agent’s mistakes?

Yes, for every alert disposition they sign, because closing or escalating an alert is a judgement the agent prepares but does not own. An AI agent can gather transaction history, draft the case narrative and recommend a disposition. The investigator who accepts that recommendation owns the decision, and the record shows exactly what they had in front of them.

Why does closing an alert stay with a person?

Anti-money laundering work rests on judgement about intent, and the FATF Recommendations build the whole system on a risk-based approach and on reporting suspicious transactions. A decision not to report is as consequential as a decision to report. Most firms therefore keep alert closure held for a named investigator, and keep the filing of any suspicious transaction report with the nominated reporting officer.

That puts the investigator at the centre of the decision chain. The agent's narrative is a draft at depth 2. The investigator's signature is what turns it into a finding.

What may the agent do, and what stays with the investigator?

• The agent may: pull account and counterparty history, summarise prior alerts, compare activity with the customer's declared profile, draft a narrative and recommend close, escalate or request more information.
• The agent should refuse: contacting the customer about the alert, since that risks tipping them off, and drawing conclusions from sources it cannot trace.
• The investigator keeps: the disposition itself, any decision to escalate to the reporting officer, and judgement about whether the customer's explanation is credible.

What does the record show about who decided?

The record is where accountability becomes concrete. For each alert, it shows every source the agent used and its provenance class. A customer's explanation of a large deposit is declared: stated, not verified. A sale contract from a solicitor is signed: a named party answers for it. The core banking ledger is sealed.

The record also names the investigator who approved the disposition, when, and whether they opened the underlying documents. If an investigator closed an alert on the strength of a declared explanation the agent presented as settled, the record shows whether the agent labelled it correctly. If it did, the closure is the investigator's. If it did not, the fault sits with the agent's design and whoever certified it.

An example: the deposit that matched the story

Take a mid-sized bank. An alert fires on three cash deposits of 9,000 each into a personal account. The agent's narrative notes the customer told a branch colleague the cash came from selling a car, and recommends closing the alert as consistent with the explanation. The source is marked declared. No sale document is on file.

The investigator, working through a large queue, accepts the recommendation. Months later the account appears in a law enforcement request. The record shows the declared label, the missing evidence and the investigator's approval in under a minute. Accountability for that closure is the investigator's, and the team lead's for allowing closures on declared sources alone.

What makes accountability fair to investigators?

Accountability is only fair if competence came first. That means:

• reviewers practise on real alerts under supervision before they close alone; a two-week practicum for AML investigators does this, and a practicum log template records it;
• a rule that a declared explanation cannot close an alert by itself, built into the agent's bounds;
• investigators stay current as the agent changes, since agent certificates lapse after 90 days and the bounds may move with each renewal; see how to prepare AML investigators for 90-day recertification.

Other roles face the same question in different forms. Underwriters approving agent-drafted quotes carry a similar weight. The broader topic sits in AI literacy and certification for people who work with agents.

Where the Academy fits

The Certified Pryme Operator credential from Pryme Intelligence Academy pairs a written exam with two weeks of real review decisions, signed off by a Pryme assessor and failed outright by any missed red line. For an AML team, closing an alert on a declared explanation alone is the kind of red line the practicum is designed to catch before it reaches a live queue.