The Sieve Agent

An AI analyst built on Claude. It works on top of the rule engine, never instead of it.

• Investigations: when a case opens, the agent reads the customer’s timeline, network and screening results and writes the summary your analyst reads first: what happened, in order, with event ids cited, what points the other way, and a recommendation with its confidence.
• Sequence hunting: every hour it reads customers who tripped rules but stayed under every pattern, and proposes new patterns. Proposals come with a backtest and can only start in Watch mode.
• STR drafts: a goAML-ready narrative for your country’s financial intelligence unit, with the deadline tracked on the case: Kenya’s FRC within two days (POCAMLA s.44), Nigeria’s NFIU within 24 hours (MLPPA 2022, CBN AML/CFT/CPF Regulations 2022). Reminders go out halfway through the window and 4 hours before it ends.
• Chat: “Why was cus_1001 blocked?” It looks things up with read-only tools before answering.

Guardrails

• Customer names, phones and emails are never sent to the model.
• Record content is wrapped as untrusted data; instructions hidden inside records are ignored.
• The agent recommends; it cannot close cases, change rules or move money.
• Without an API key, cases still get a summary assembled from rule reasons.