AI Prototype for Financial Compliance: Application of Multi-Agent Workflow in Sanction Screening

This article introduces the Agentic AI prototype project FS_kyc_aml_alert_triage in the financial services sector, demonstrating how to use a multi-agent workflow to assist compliance analysts in reviewing OFAC-style sanction alerts, evaluating evidence, and generating auditable disposition recommendations. The core goal of the project is not to replace analysts, but to structure the investigation process, apply conservative decision-making logic, and generate compliance justifications. As an educational demonstration prototype, it requires controlled testing before practical application.

Practical Challenges of Sanction Screening

• Massive alerts: Large banks receive thousands to tens of thousands of alerts daily
• High false positive rate: Fuzzy matching and complex scenarios lead to a false positive rate exceeding 90%
• Compliance pressure: Missed reports may lead to regulatory penalties and reputational damage
• Audit requirements: Disposition decisions need clear basis and complete trails

Project Positioning

The core goal is to demonstrate how Agentic AI can structure investigations, make conservative decisions, and generate compliance justifications. It is clearly an educational demonstration prototype, not a production-level system.

Multi-Agent Workflow

Four-stage pipeline: Alert candidate → Name Matching Agent → Context Agent → Justification Agent → Recommended Disposition

• Name Matching Agent: Uses rapidfuzz for fuzzy matching, outputs matching strength and reasons to ensure interpretability
• Context Agent: Integrates background information such as counterparty, transaction amount frequency, and geographic location
• Justification Agent: Synthesizes outputs from the previous two stages to form a matching summary, risk analysis, mitigating factors, and uncertainty explanations
• Disposition Recommendation: Options include approve/ escalate/ block; boundary cases tend to be escalated

Tech Stack

• LangGraph: State management, conditional branching, observability, human-machine collaboration
• Hybrid strategy: Rule engine (name matching) + LLM (context analysis) + structured output

Synthetic Dataset

• Design purpose: Simulate real scenarios under privacy protection
• Covered scenarios: True positives, false positives, boundary cases
• Scale: Small proof of concept

Evaluation Methods

• Expected disposition verification: Consistency between output and expectations
• Justification quality assessment: Clear and accurate explanations
• Boundary case testing: Handling of ambiguous situations

Disclaimer: Prototype verification, not regulatory-level model validation

Key Design Principles

1. Auditability: Each recommendation includes a complete reasoning chain (matching parameters, risk weights, LLM justifications, confidence level)
2. Conservative Decision-Making: Escalate in ambiguous cases, supplement missing information, and introduce manual review for complex scenarios
3. Human-Machine Collaboration: AI provides analytical recommendations, humans make final decisions, and overriding AI recommendations is supported
4. Modular Improvement: Components can be optimized independently (e.g., the name matching algorithm can be upgraded separately)

Application Scenarios

1. L1 Analyst Support: Structured investigation checklists, information aggregation, preliminary assessment, and case references
2. Training Tool: Demonstrates standard processes, provides case materials, and helps new analysts build intuition
3. Process Standardization: Unifies checkpoints, reduces fluctuations due to experience differences, and establishes processing benchmarks

Limitations

• Synthetic data: May differ from real data distribution
• Preset matching: Does not cover the matching discovery phase
• Non-official lists: Not connected to real OFAC SDN lists
• Output parsing: LLM output text parsing instead of strict schema validation
• Evaluation scope: Prototype-level verification

Usage Notes

• Educational demonstration project, not legal/compliance advice
• Production deployment requires additional controls, testing, and governance
• Final decision responsibility lies with human analysts and institutions

Future Development Directions

• Connect to real sanction list data sources
• Expand to other compliance screening types such as anti-money laundering (AML)
• Integrate more external data sources to enhance context
• Fine-grained confidence calibration
• Production-level monitoring feedback loop

Industry Insights

Agentic AI has potential in highly regulated industries: structuring complex tasks, maintaining human oversight, and generating auditable records; design needs to be conservative, transparent, and human-machine collaborative.

Conclusion: This project provides a reference for the application of Agentic AI in financial compliance, demonstrating the possibility of AI and human collaboration to address compliance challenges. Production systems require more engineering and regulatory reviews.