← All benchmarks

AML Bench

Professional Work · 2026-10-19

Veris AI's anti-money-laundering agent benchmark: the agent is the second-level investigator at a simulated mid-sized US bank, handed 153 alerts by a first-level colleague who stays on the chat. It pulls ledgers and KYC records from an Apache Fineract core-banking twin, re-screens the parties against a sanctions and PEP watchlist, applies the bank's written procedure and files a structured report, narrative and case action in a Marble case-management twin. In 48 cases the deciding record is held by the colleague and must be asked for. An attempt passes only if every grading criterion passes; grading reads the state the agent leaves in the case system, not its chat (the page's Dataset metadata says an LLM judge grades from the recorded service requests). Every agent runs each case three times, and pass@1 is the share of passing attempts. Rows are a model in a named coding-agent harness, with no reasoning effort stated.

Top models (higher is better)