Can an AI agent investigate an AML case?

It can run most of the investigation, but not reach the finding. An AI agent can assemble the case file, map the flow of funds, compare activity with the customer's profile and draft a findings note with options. Whether there is suspicion, whether to report and what happens to the customer stay with the investigator and the reporting officer.

What does an AML investigation involve?

An investigation begins when an alert, a referral or a tip is escalated beyond first review. The investigator's job is to understand what happened, whether there is an innocent explanation, and whether the facts give grounds for suspicion. The output is a decision: close, keep monitoring, or refer to the reporting officer with a recommendation to report.

Most of an investigator's time goes on gathering. Statements, KYC files, prior alerts, counterparty details, open-source information and internal notes all have to be pulled and put in order before any thinking can start. A complex case can involve dozens of accounts and hundreds of transactions. That gathering and ordering is what the agent does.

What may the agent prepare?

The agent may:

• Build the chronology: every relevant transaction, contact, alert and change to the customer record, in date order.
• Map funds: where money came from and went, with totals by counterparty, and second-level flows where its access allows.
• Compare the activity with the customer's stated purpose, occupation and expected activity.
• Pull prior alerts, cases and reports on the customer and connected parties.
• Run adverse media and open-source searches and summarise relevant results with sources.
• Draft questions for the customer, for the investigator to approve and send through the normal channel.
• Draft the findings note: facts, red flags against known typologies, possible innocent explanations, and options.

How the file is laid out inside your systems is covered in how to let an AI agent build case files in your case management system.

The investigator and reporting officer decide:

• What the facts mean, and whether an innocent explanation holds.
• Whether there are grounds for suspicion.
• Whether to refer, report, and in what terms.
• Whether to restrict, exit or keep the relationship.
• Whether and how to contact the customer.

What does an agent-prepared case look like?

Take a business bank. An alert on a construction subcontractor's account is escalated: large inbound payments from three main contractors, followed by transfers to fifteen personal accounts each week.

The agent builds the file. The chronology covers six months. The funds map shows the fifteen personal accounts receive similar sums each week, consistent with wages. But the agent also finds that eight of the fifteen accounts are at other banks and opened within the same month, that four share an address, and that the subcontractor's accounts show no tax or pension payments at all. Adverse media turns up nothing on the company. A registry check shows it was incorporated nine months ago.

The draft findings note sets out two readings. The innocent reading is a new company paying cash-in-hand workers, which is a tax issue but not necessarily laundering. The concerning reading is a labour exploitation pattern, with wages paid into accounts controlled by others. The note lists what would tell them apart: whether the workers' accounts show ordinary spending or immediate onward transfers.

The investigator asks for that analysis on the four accounts at the bank's own customers. Three show immediate transfers out to a single account. She refers to the reporting officer, who files. The decision path is described further in how to prepare AML case investigations with an AI agent.

Which controls keep the investigation sound?

• Both readings drafted. The agent's note sets out evidence against suspicion as well as for it.
• Every fact sourced. Each statement links to a record, and inferred links are labelled as inferred.
• No contact without approval. Customer questions are drafted, approved by the investigator, and sent through normal channels.
• Restricted access. Investigation files are visible only to the team.
• Decisions recorded with reasons, including decisions to close.

Where your firm operates in the EU, consider how the EU AI Act treats agents that assess risk about people; see EU AI Act and AI agents in compliance and AML teams. The general principle is in AI compliance agents: second line, never the last word.

Where Pryme fits

In Pryme, building the chronology, tracing funds and drafting findings are separate capabilities, each certified on your own closed cases. The agent can read the case management system but cannot close a case or contact the customer. Every fact in the findings note carries its provenance, so the investigator can weigh ledger records above open-source claims.