Alert triage without closing alerts

An agent can do most of the work on an AML alert except the last step. It reads the rule that fired, gathers the customer's profile and history, analyses the transactions, checks open sources and drafts a cited disposition narrative. An analyst reads the evidence and decides whether to close or escalate. Reporting stays with the nominated officer.

What is AML alert triage, and who owns it?

Transaction monitoring systems raise alerts when activity matches a scenario: cash deposits just under a reporting threshold, rapid movement through an account, payments to higher-risk places. Most alerts turn out to have an innocent explanation. Each still needs a person to look, decide and record why. Alert triage is that first review.

First-line or financial crime operations teams own the queue. The nominated officer, often the MLRO, owns the decision to report suspicion. Under the FATF Recommendations, the international standard for AML, firms take a risk-based approach and must report suspicious transactions. Both duties rest on human judgement that can be explained later. That is why our view, set out in AI compliance agents: second line, never the last word, is that the agent prepares and people decide.

What are the steps?

1. Agent, depth 1: reads the alert. The scenario that fired, its parameters, the transactions that triggered it. It cites the rule by its identifier, so the analyst sees exactly what was being tested.
2. Agent, depth 1: gathers customer context. Due diligence profile, stated purpose and expected activity, risk rating, previous alerts and how they were closed.
3. Agent, depth 1: analyses the transactions. Counterparties, places, amounts, timing, and how the activity compares with the customer's own history and stated purpose.
4. Agent, depth 1: checks open sources. Adverse media on the customer and key counterparties, labelled as external.
5. Agent, depth 2: drafts the disposition. Facts first, each cited. Then which indicators of the scenario are present and which are absent. Then innocent explanations it checked and what it found. Then open questions. A proposed outcome comes last, labelled as a proposal.
6. Agent, depth 1: ranks the queue. By risk, using rules the financial crime team sets, so the analyst sees the most important alerts first.
7. Person, analyst: decides. Closes the alert with a reason, or escalates it for investigation.
8. Person, nominated officer: decides on reporting for escalated cases.

Why does closure stay with a person?

Closing an alert is a decision that the activity is not suspicious. If that decision is wrong, the firm has failed to report. Someone must be able to explain, later and to a supervisor, why they looked at this evidence and reached that view. Whether an agent can close AML alerts is answered no for that reason.

There is a practical reason too. The value of a person in the loop disappears if they agree with every proposal without reading it. Keeping closure with the analyst, and making the analyst engage with evidence before they see a proposal, is what keeps the person's judgement real.

Which controls apply, and what evidence do they leave?

Evidence before proposal. The draft puts facts and indicators before the proposed outcome, so the analyst forms a view first. Evidence: the draft structure, fixed in the template.

Rule cited. Every disposition names the scenario and parameters that fired. Evidence: the rule identifier in the record, which also supports later tuning.

Provenance on every fact. Your own transaction data is sealed. The customer's stated purpose is declared. Adverse media is external. The analyst sees which is which.

Agreement tracking. How often analysts accept the agent's proposal unchanged, by analyst and by scenario. Very high agreement deserves a closer look.

Closure by a named person. The analyst's identity, decision and reason are filed with the draft. The AML alert disposition record template lists the fields.

What should you measure?

• Time from alert to decision.
• Proposals accepted unchanged, by analyst and scenario.
• Quality assurance findings on agent-prepared alerts, compared with those prepared by people alone.
• Escalated alerts that led to reports. Shows whether escalation proposals are well aimed.
• Aged alerts. Backlogs are often why firms start here.

Where does it usually go wrong?

Automation bias. Analysts start trusting the proposal and skimming the evidence. Agreement rates rise and quality assurance findings follow. Watch both.

Bulk closure. Many similar alerts with similar proposals invite closing them together. Each alert still needs its own decision.

Thin context on repeat alerts. A customer alerting for the fifth time gets the same narrative as the first. The draft should say what is new since the last disposition, and show earlier closures together, because a run of innocent explanations can itself be a reason to look harder.

Tuning on proposals. Using the agent's proposed outcomes to tune scenarios builds its view into the rules. Tune on analysts' decisions, not the agent's.

Sanctions mixed in. Sanctions screening alerts follow different rules and deadlines; see triaging sanctions screening alerts and whether an agent can clear a sanctions alert.