How We Built the Agentic AI-Based Scenario Developer Assistant for Anti-Money Laundering Solutions

The result is Aurora (AML Unified Rule Orchestration & Risk Agent), a solution that creates a bridge between human intent and automated action. We are proud that Aurora didn’t just work—it resonated. Our team took home awards in 3 separate categories: Agentic AI & Decisioning, Trustworthy AI, and Banking.

Anyone in Compliance knows the struggle. A new risk emerges, and a Compliance Officer writes a rule. But turning that rule into live code involves data teams, endless meetings, manual testing, and complex validation. This process can be slow, prone to misinterpretation, and expensive. That's where Aurora steps in. It speeds up several parts of this process with its agents and tools:

• transforming business requirements into technical implementation using a Domain-Specific Language;
• mapping business terms to table columns;
• generating test data for high coverage white-box testing;
• checking business requirement completeness and consistency;
• validating scenario and data consistency;
• running test cases;
• comparing test results with expected results; and generating documentation.

Using SAS Intelligent Decisioning, Visual Investigator (VI), and Large Language Models (LLMs), we built a system where a user can describe a scenario in natural language, and Aurora handles the rest. Aurora represents a practical implementation of SAS' Agentic AI vision. It is an AI that doesn’t just respond to text; it acts on it—while remaining fully governed, explainable, and human-controlled.

How does it work?

A Team of Specialized Agents Aurora isn't a single model; it’s an orchestrated workflow of specialized agents, each with a distinct role. E.g.:

• Rule Engineer: Converts free text into formal, domain-specific AML scenario definitions.
• Completeness Checker: Validates whether the request has enough detail or if clarification is needed.
• Governance Agent: It checks for overlaps with existing scenarios and scans for potential bias.
• Recalibrator: Tests and retrains the alert-scoring model using historical data to ensure fairness and accuracy.

We built Aurora with a Deterministic Agentic AI architecture. You cannot just let an LLM hallucinate code or tasks. That is why we implemented several guardrails on top of a deterministic Agent Workflow.

• “Jailbreak” Guard: Checks for possible manipulations.
• No Arbitrary Code: Users cannot execute random commands; the system only uses pre-defined, safe tools within SAS Viya.
• Human-in-the-Loop: Aurora automates the work, not the authority. Before any scenario goes live, a human compliance officer must review and approve the artifacts.

By integrating GenAI at well-defined points within the robust SAS Viya environment, we can achieve:

• Speed: Drastically reduced time-to-market for new AML rules.
• Efficiency: Automated unit testing and documentation generation.
• Accuracy: Real-time model recalibration to reduce false positives.

Aurora shows an approach to implement an Agentic AI use casein banking that allows humans and machines to work side-by-side.