Anti-Money Laundering Multi-agent Orchestration with AWS Strands

Global Impact of Anti-Money Laundering

• Anti-money laundering (AML) is a global crisis, with $3.1 trillion in illicit funds flowing in 2023 and similar patterns in 2024
• Majority of these funds come from drug trafficking, human trafficking, terrorism, and fraud
• Financial institutions have been working on AML regulations for a long time, but no single solution can solve this problem alone

Evolution of AML Standards

• Started with the Bank Secrecy Act, followed by the creation of FATF with 40 specific standards
• 1992 saw the introduction of Suspicious Activity Reports (SARs) for any deviations in financial transactions
• 2020 brought the Anti-Money Laundering Act, further increasing the responsibility and vigilance for financial institutions

Challenges with Traditional AML Approaches

• Financial institutions deal with diverse data sources, including financial transactions, customer profiles, KYC data, and news feeds
• High volume and velocity of data overwhelms traditional rule-based systems, leading to up to 95% false alarms
• This slows down the detection of real threats and wastes resources on reviewing false positives

Agentic AI Workflow for AML

Agent-based Approach

• Agents are autonomous entities that make decisions and take actions to accomplish goals with minimal human intervention
• Agent workflows involve the coordinated sequence of multiple agents working together to solve complex problems

Key Technologies

• AWS Strands SDK: Open-source framework for easily creating agents
• Amazon Bedrock Agent Core: Composite suite of services for deploying agents at scale, including runtime, gateway, memory, and observability

Agent Workflow Components

1. Perception Agent: Understands the transaction and collects relevant data from various sources
2. Context Agent: Retrieves historical transaction data and KYC information to provide context
3. Reasoning Agent: Leverages the collected data to assess the risk of the transaction
4. Risk Agent: Integrates with existing risk models to provide a comprehensive risk score
5. Action Agent: Automates actions based on the risk score, such as straight-through processing or filing SARs
6. Audit Agent: Records the entire decision-making process for compliance and auditing purposes
7. Learning Agent: Continuously updates the system based on the outcomes of past transactions

Technical Implementation Details

• Perception Agent uses Kinesis to process real-time transaction data and connects to various data sources
• Context Agent retrieves historical data from DynamoDB and uses Lambda functions and the Agent Core Gateway to access external data
• Reasoning Agent uses the Strands SDK to make decisions based on the collected data and context
• Risk Agent integrates with existing risk models, either through APIs or by invoking Lambda functions
• Action Agent automates actions like notifications, transaction flagging, and SAR generation
• Audit Agent logs all agent interactions and decisions in Elasticsearch for compliance and auditing
• Learning Agent updates the system's knowledge base in DynamoDB to improve future decision-making

Business Impact and Use Cases

• Reduces manual review and false positives by automating the AML process
• Accelerates the detection of suspicious transactions and the generation of SARs
• Provides a comprehensive audit trail of the decision-making process for compliance
• Enables proactive risk management by incorporating real-time news and external data
• Improves the overall efficiency and effectiveness of AML efforts within financial institutions

Demonstration Scenarios

1. Behavioral Signal Triage: Identifies a suspicious transaction involving a high-risk sender and receiver, resulting in the recommendation to file an SAR and freeze the transaction.
2. Repetition Pulse Scan: Incorporates real-time news and sanctions data to detect a transaction involving a sanctioned individual, leading to the decision to file an SAR and monitor the transaction.
3. Complex Orchestration: Combines the risk assessment and learning components to handle a low-risk payroll transaction, demonstrating the system's ability to adapt to different scenarios.

Key Takeaways

• The agentic AI workflow leverages the flexibility of the Strands SDK and the scalability of the Bedrock Agent Core to build a modular, extensible, and highly automated AML solution.
• By integrating diverse data sources, applying reasoning and risk assessment, and automating actions, the system significantly improves the efficiency and effectiveness of AML efforts.
• The comprehensive audit trail and continuous learning capabilities ensure the system evolves to address emerging threats and maintain compliance.
• The demonstrated use cases showcase the versatility of the solution in handling various AML scenarios, from straightforward transactions to complex, high-risk situations.