Keep one line fixed: the agent drafts, the nominated officer decides and files. Within that line, the agent can assemble the subjects, the transactions, the timeline and the grounds for suspicion from the investigation file, and write a clear narrative against your reporting standard. The MLRO or nominated officer reviews, decides whether to report, and submits.
What is the process, and who owns it?
A suspicious activity report tells the national financial intelligence unit why the firm suspects money laundering or terrorist financing. FATF Recommendations set suspicious transaction reporting as a core obligation, and each country's law decides the form, the threshold of suspicion and the deadlines. A good report is specific: who, what, when, how much, and why it is suspicious. A vague one is of little use to the authorities and does little to protect the firm.
The nominated officer, usually the MLRO, owns the decision to report and the submission. Investigators prepare the case. In most firms the narrative is the slowest part, and quality varies by who wrote it. That is the work the agent takes. The reason it never takes the filing is set out in an agent can draft a suspicious activity report but never file one.
What are the steps?
- Person, investigator: completes the investigation and recommends reporting, or not, with reasons.
- Agent, depth 1: reads the case file. Alerts, transactions, customer due diligence, prior reports, open-source findings and investigator notes.
- Agent, depth 1: extracts the reportable facts. Subjects with identifiers, accounts, counterparties, transactions with dates and amounts, and the period covered. Each fact cites its source in the file.
- Agent, depth 2: drafts the narrative. What the activity is, why it is unusual for this customer, which indicators apply, what the firm did, and what it does not know. Plain language, chronological where possible.
- Agent, depth 1: checks against the drafting standard in your SAR drafting checklist: all fields complete, amounts reconciled, no speculation presented as fact, no tipping-off risk in any customer communication linked to the case.
- Person, nominated officer: reviews. Decides whether suspicion is established, edits the narrative, and decides on consent requests or other immediate steps where the law provides for them.
- Person, nominated officer: submits to the financial intelligence unit and records the reference.
- Agent, depth 1: updates the case with the submission reference and sets follow-up dates. A person confirms.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Investigation | Nothing | None | Investigator's recommendation |
| Read file | Reads everything in the case | 1 | None |
| Facts | Extracts with source references | 1 | Correcting errors |
| Narrative | Drafts against the standard | 2 | The final wording |
| Checklist | Tests completeness and consistency | 1 | None |
| Decision | Nothing | None | Whether to report |
| Submission | Nothing | None | Filing |
| Case update | Records reference and dates | 1 | Confirming |
Which controls apply, and what evidence does each leave?
No filing capability. The agent has no access to the reporting portal. Evidence: the access register.
Facts traced to the file. Every fact in the draft cites its source document or record. Evidence: the nominated officer can check any line, and so can an inspector later.
Narrative separated from decision. The draft does not say whether to report; it sets out the facts and indicators. The decision is recorded separately. Evidence: a decision record signed by the nominated officer, including decisions not to report.
Tipping-off guard. The agent drafting the report must not also be the agent talking to the customer. Evidence: role separation in the agent inventory and access configuration.
Draft and final kept. Evidence: the officer's edits, which over time show where drafts are weak.
What should you measure?
- Time from investigation close to submission.
- Officer edits per draft, by section.
- Facts corrected at review, especially amounts and dates.
- Feedback from the financial intelligence unit, where it gives any, on report quality.
- Reports with incomplete subject identifiers.
- Drafts the officer decided not to file, and why. A high number may mean investigations escalate too readily, or that the drafting standard pushes weak cases forward.
Where does it usually go wrong?
Drafts that argue for suspicion. An agent asked to write a SAR will tend to make the case. The narrative should be factual and balanced, including the explanations the firm considered and why they did not dispel suspicion.
Unknowns left out. A narrative that reads as complete when the firm could not trace where funds went, or could not reach the customer, overstates what is known. The draft should have a short section on what the firm does not know and why, because that tells the authorities where their own enquiries might start.
Copying the alert text. The alert explains why a rule fired, not why the activity is suspicious. Reports that restate rule logic are weak. The narrative should describe the customer's behaviour in plain terms, so a reader with no knowledge of your monitoring system understands why it matters.
Errors in identifiers. A wrong date of birth or account number makes a report harder to use. The checklist should reconcile every identifier against the customer record.
Agent used to decide the threshold. Whether suspicion is established is a legal judgement for the nominated officer. Whether an AI agent can file a suspicious activity report explains why this remains a person's role.
Policy drift. When reporting guidance changes, the drafting standard must change too. Run it through the same process as drafting policy updates for approval.
The wider stance is in AI compliance agents.
Questions people ask
Can the agent recommend whether to report?
It can list the indicators present and absent. A recommendation belongs to the investigator, and the decision to the nominated officer, and both are recorded.
Does the agent need access to previous reports?
It helps, for continuity and for referencing earlier submissions on the same subject. Access to prior reports should be tightly restricted and logged, because they are highly sensitive.
Can the agent draft reports in bulk for similar cases?
It can draft each one, but each must be reviewed on its own facts. Batch-similar narratives are a warning sign to any reader of the reports.