No, but it can prepare most of them for closing. An AI agent can review each transaction monitoring alert, pull the customer's history, test the activity against what the firm knows, and draft a closure rationale. Closing the alert is a judgement that there is no suspicion, and it stays with a trained analyst who signs it.
Why does closure need a person?
Closing an AML alert is a decision with legal weight. It says the firm looked at activity its own monitoring flagged and concluded it was not suspicious. If that conclusion was wrong, the firm may have failed to report, which supervisors treat seriously. They also examine closure quality directly: whether rationales are specific, whether the evidence supports them, and whether analysts are closing alerts too quickly.
An agent cannot carry that accountability. There is also a practical risk. If an agent closes alerts, the firm's view of its own risk is filtered through a judgement no person checked, and errors repeat at volume before anyone sees them.
What may the agent do, and what stays with the analyst?
| Task | The agent does | Depth | Stays with the analyst |
|---|---|---|---|
| Alert review | Reads the alert, the scenario that fired and the transactions | 1 | Nothing |
| Customer context | Pulls the profile, expected activity, past alerts, KYC refresh date | 1 | Nothing |
| Analysis | Compares activity with expected behaviour; looks for linked parties | 1 to 2 | Interpreting what it means |
| Rationale | Drafts a specific closure or escalation rationale with evidence | 2 | Accepting, editing or rejecting |
| Closure | Never | None | Closing the alert and signing |
| Escalation | Prepares the case file for investigation | 2 | Deciding to open a case |
| Feedback | Logs patterns in false positives for scenario tuning | 2 | Changing scenarios or thresholds |
Moving from alert to full investigation is covered in whether an AI agent can investigate an AML case.
What about obvious false positives?
Every firm has them, sometimes in large numbers. The answer is not to let the agent close them on judgement. It is to fix the scenarios that generate them, through tuning approved by compliance and tested on past data. A threshold change agreed by people is a controlled decision; an agent quietly closing a category of alerts is not.
The agent helps by spotting the patterns: "most alerts from this scenario involve payroll payments from a known employer". Compliance decides whether to tune.
What does a good closure rationale look like?
Specific, sourced and honest about what was not checked. "Activity consistent with profile" is not a rationale. A good one says what fired, what the agent compared it with, what explained it, and what it could not verify.
For example: "Alert fired on three cash deposits totalling 9,000 over five days. Customer's profile lists self-employment as a market trader; deposits on market days match six months of prior pattern. No linked accounts. Source of funds not independently verified." The analyst can test each statement quickly.
An illustrative example
Take a mid-sized bank whose monitoring system generates a few hundred alerts a week. The agent prepares each one before an analyst opens it.
On one alert, a dormant personal account suddenly receives several transfers from different people, followed by a single transfer out. The agent pulls the history, notes the account was inactive for two years, finds that two of the senders have accounts at the bank that were themselves flagged last quarter, and drafts an escalation rationale. The analyst reads it, agrees, and opens an investigation. On another, a business customer's monthly turnover spikes. The agent finds the spike matches a seasonal pattern in the previous two years and drafts a closure rationale. The analyst checks the prior years, agrees, and closes the alert with her name on it.
What do the rules say?
FATF's risk-based approach expects firms to monitor transactions and report suspicion, with effective human judgement in the process; see FATF Recommendations and AI agents in AML. In the EU, the AI Act adds obligations depending on how systems are used, covered in the EU AI Act and AI agents in compliance and AML teams. Neither suggests that closure can be handed to software.
The same pattern applies to other alert types, such as trade and communications surveillance; see how to review trade surveillance alerts with an AI agent and how to review communications surveillance alerts with an AI agent.
What controls keep this safe?
- No closure capability for the agent, enforced at the moment of action.
- Rationales tested against past alerts with known outcomes, including ones that became reports.
- Analyst decisions sampled by quality assurance, with attention to closures that followed the agent's draft unchanged.
- Scenario tuning kept separate and approved by compliance.
- Every alert, draft and decision filed.
The wider principle is in AI compliance agents: second line, never the last word.
Questions people ask
Does the agent reduce alert backlogs?
Yes, by cutting the time each alert takes. Analysts still close every one, but faster and with better rationales.
Should the agent's draft be visible to supervisors?
The firm's records should show both the draft and the analyst's final decision, so reviewers can see where the analyst agreed or changed it.
What if analysts disagree with the agent often?
That is useful information. It may mean the agent needs better sources or a narrower brief, or that analysts are catching things it misses.