An agent can draft a suspicious activity report but never file one

An AI agent can gather the facts, build the timeline and draft a suspicious activity report for review. It should never file one. Filing records a judgement that suspicion exists, made by a person who answers for it, and it is a legal act with consequences for the customer and the firm. That act stays with the responsible person.

What can the agent usefully do in SAR work?

A great deal, and most of it is the slow part. Investigators spend much of their time assembling material: pulling transactions, reading account notes, checking counterparties, putting events in order and writing the narrative. An agent can do that preparation well, provided it works from the firm's own records and states where each fact came from.

At depth 1 it can read the alert, the customer file and the transaction history and summarise what is there. At depth 2 it can draft a suspicious activity report for review: a structured narrative of who, what, when, where and why the activity looks unusual, with each statement linked to a source record. The investigator starts from a coherent draft rather than a blank page and a dozen open screens.

Why must filing stay with a person?

A suspicious activity report is not a summary of data. It is a statement, made to an authority, that someone in the firm knows or suspects something. The FATF Recommendations set suspicious transaction reporting as a core obligation of the international anti-money-laundering standard, and national rules generally place the decision on a responsible person, often a nominated officer, rather than on a system.

Three things make filing a human act:

1. It is a judgement. Whether facts amount to suspicion depends on context, experience and the firm's knowledge of the customer. Two investigators can read the same draft and reach different conclusions; one of them has to decide.
2. It carries accountability. If a report should have been filed and was not, or was filed carelessly, someone must answer for it. An agent cannot.
3. It has consequences. A filed report can affect the customer's relationship with the firm and with other institutions, and it triggers obligations for the firm, including keeping the report confidential.

The short answer to whether an AI agent can file a suspicious activity report is no, and the reason is not that the technology could not submit the form. It is that the act needs someone who can be held to it.

Isn't the draft itself a kind of decision?

Partly, and this is the subtle risk. A draft frames the facts. If the agent writes a persuasive narrative of suspicion, the investigator may anchor on it and approve without testing it. If the agent writes a bland one, real concerns may be underplayed. Either way, the draft can shape the outcome more than the person realises.

The controls for this are practical:

• The draft should separate observed facts from the agent's interpretation, so the investigator can see where the agent is inferring.
• It should include facts that point away from suspicion as well as towards it, such as a known business reason for an unusual payment.
• It should record what the agent looked for and did not find, not only what it found.
• Every factual statement should cite the record it came from, with the record's provenance.

A suspicious activity report drafting checklist turns these into checks an investigator can run quickly.

How should the boundary be enforced?

Not by instruction alone. The agent's role should simply not include the capability to submit reports, so any attempt is refused at the moment of action. Within the firm's own limits, the drafting capability can run at depth 2, and the filing step sits outside the agent's bounds entirely.

This is a narrow example of a wider pattern. The discussion of GDPR Article 22 and agents that draft but do not decide explains why keeping decisions with people matters where outcomes have significant effects on individuals.

What about the decision not to file?

It is just as much a judgement, and it deserves just as much evidence. Many investigations end without a report. The record should show what was reviewed, the reasoning for concluding that no report was needed, and who made that call. An agent can draft the closure note in the same way it drafts a report. The person signs it.

Firms often record "no further action" thinly. If an agent is doing the preparation, there is no longer a good excuse for that: the same structured material that supports a filing supports a reasoned decision not to file.

What about tipping off?

Any agent involved in suspicious activity work must be kept away from customer contact on the same matter, and customer-facing agents must not be able to see that a report is being considered or has been made. Disclosing either is a serious offence in many jurisdictions. The safest design is that the drafting agent and any customer-facing agent share no sources about the investigation at all.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

Where Pryme fits

In Pryme, a drafting agent can be certified at depth 2 for SAR preparation on the firm's own past cases, with each statement carrying its provenance class. Submission is outside the agent's role, so any attempt is refused and filed as a verdict. This sits within the wider approach to AI compliance agents as second line, never the last word.