Inside the First Multi-Agent AI Pilot for Banking Fraud and AML: A 29% Detection Improvement

Consortium Pilot: 10 Global Banks Slash False Positives with Multi-Agent AI As of May 27, 2026, a consortium of 10 global banks has quietly rewritten the playbook for financial crime compliance. In the first documented pilot of its kind, a multi-agent AI system built on Amazon Bedrock slashed false positives by 24% while surfacing 29% more genuinely suspicious activity than legacy rules-based systems. The results, shared with industry observers under a collaborative research agreement, offer the most concrete evidence yet that agentic AI can meet the twin demands of regulatory rigor and operational efficiency. Banks have long been stuck in a compliance trap: tightening anti-money laundering (AML) and fraud controls drives up alert volumes, but most alerts turn out to be innocent. Investigators drown in noise while sophisticated criminals slip through. The consortium—comprising ten instit

utions from North America, Europe, and Asia, none publicly named due to competitive sensitivity—set out to test whether specialized AI agents working in concert could break the cycle. Their findings present a vendor-neutral, replicable blueprint that B2B operations leaders can use to evaluate multi-agent systems for high-compliance environments. The Consortium Pilot: Why 10 Banks Joined Forces for AI The initiative, which kicked off in late 2025, was born from a shared frustration. Despite heavy investment in transaction monitoring, watchlist screening, and case management, individual banks faced an 85–95% false positive rate on automated alerts, while regulatory penalties for missed suspicious activity reports (SARs) continued to climb. The Financial Action Task Force (FATF) had flagged “inefficient alert management” as a systemic weakness in its 2024 guidance, and the US Office of the

Comptroller of the Currency (OCC) was increasingly critical of banks that failed to modernize compliance tooling. Ten banks agreed to pool anonymized data and design a joint pilot under a pre-competitive data trust—a model already proven in areas like cybersecurity threat intelligence. Their goals were ambitious: use agentic AI to triage alerts more accurately, reduce investigation time per case, and ensure full explainability for examiners. The pilot ran on a shared AWS environment, with each bank contributing labeled alert data but no raw customer information, preserving privacy while creating a robust training and testing corpus. Multi-Agent Architecture on AWS Bedrock: Blueprint Overview The architecture is a textbook example of how to decompose a complex compliance workflow into specialized, collaborating agents. The system was built using Amazon Bedrock’s multi-agent collaboration

capabilities (generally available as of early 2026) and two open-weight large language models: Anthropic’s Claude 5 Haiku and Meta’s Llama 5, chosen for their balance of speed, cost, and reasoning depth. Orchestration layer: A central supervisor agent (using Claude 5 Haiku) interprets incoming transaction alerts, dispatches tasks to subordinate agents, and synthesizes their outputs into a human-readable case summary. This design avoids a single monolithic model struggling with context retention across multiple data sources. Data flow: 1. Ingestion feeds (SWIFT messages, ACH batches, trade logs) hit a streaming pipeline on AWS Kinesis. 2. The supervisor agent extracts relevant snippets and routes them to specialized agents via the Bedrock AgentCore interface. 3. Each specialized agent runs its analysis concurrently, returning structured findings within predefined confidence thresholds. 4.

The supervisor compiles a unified alert score and, if necessary, a draft SAR narrative for human review. All model calls are logged at the prompt and response level, creating a complete audit trail. As of May 2026, AWS Bedrock pricing for Claude 5 Haiku stood at $0.00125 per 1,000 input tokens and $0.005 per 1,000 output tokens, while Llama 5 was priced at $0.0008 per 1,000 input tokens and $0.0025 per 1,000 output tokens on Bedrock’s on-demand tier—cost profiles that kept the pilot economic even for high-volume alert streams. Specialized Agents: Transaction Monitoring, Watchlist Screening, and Case Prioritization The real innovation lies in how three distinct agent types cooperated, each trained on a curated dataset from consortium members and fine-tuned using supervised examples of confirmed suspicious activity. Transaction Monitoring Agent (Llama 5) This agent analyzes payment patter

ns, velocity checks, and counterparty risk. Unlike static rules, it learns to detect subtle anomalies—such as layered structuring over multiple weeks or micro-transactions below standard reporting thresholds—by considering entire account histories rather than isolated events. The pilot’s monitoring