While fraud detection models are increasingly accurate, fraud response often remains slow, manual, and fragmented. Leveraging SAS Viya's decisioning capabilities and Large Language Models (LLMs), the team built an AI powered voice agent named Agatha, to proactively contact customers when a suspicious card transaction is detected. This agentic AI use case built with SAS Viya demonstrates how analytics can move beyond detection to support real-time, customer-centric fraud resolution.
Why fraud response is the real bottleneck
Most banks already rely on analytical models and business rules to detect suspicious transactions. However, once an alert is raised, the downstream response is often inefficient. Customers may receive an SMS or app notification that goes unnoticed, or they are asked to contact a call center to resolve the issue. This introduces latency, increases operational costs, and negatively impacts customer experience—especially when legitimate payments are declined.
The Minority Reporters project focused explicitly on this "last mile" of fraud management: transforming fraud alerts into timely, actionable resolution steps through direct customer interaction.
Solution overview: Agatha, an AI voice agent
Agatha is triggered immediately when a fraud detection system flags a transaction as suspicious. Rather than sending a passive notification, Agatha initiates a phone call to the cardholder. During the call, the agent asks a small number of targeted questions designed to quickly assess whether the transaction is legitimate or fraudulent.
Based on the conversation, Agatha can support several outcomes:
- Approving the transaction if it is legitimate
- Blocking or freezing the card when fraud is confirmed
- Resetting digital banking credentials if compromise is suspected
- Escalating the case to a human operator when confidence is insufficient
The key principle is that customer interaction becomes part of the decision process rather than a disconnected follow-up step.
Architectural building blocks
The solution is structured around loosely coupled components orchestrated through SAS Viya:
- Fraud signal ingestion: A fraud signal is received from an upstream detection engine. The project deliberately focused on integration patterns rather than the internal mechanics of fraud models.
- Speech-to-text: Customer responses during the phone call are converted into text using speech-to-text services. This transformation enables subsequent natural language processing and reasoning.
- LLM-based conversation handling: A LLM interprets customer intent and generates concise, controlled responses. The LLM assists with understanding and interaction but does not directly execute business actions.
- Decision orchestration with SAS Viya: Customer intent and contextual information are passed to SAS Viya decision logic, where business rules, analytical models, and thresholds determine the final action. This ensures explainability, governance, and compliance.
- Monitoring with SAS Visual Analytics: Operational dashboards built in SAS Visual Analytics provide visibility into calls, outcomes, escalation paths, and performance metrics.
Why this architecture matters
A key takeaway from the project is the importance of separating AI reasoning from execution. LLMs enhance interaction and interpretation, while SAS Viya remains responsible for decision governance, transparency, and auditability—an essential requirement in regulated industries such as financial services.
Implementing similar solutions with the SAS Agentic AI Accelerator
For organizations interested in building similar integrations between LLMs and SAS Viya, the SAS Agentic AI Accelerator provides a structured, enterprise-ready framework.
The accelerator is designed to help teams embed LLM capabilities into governed workflows rather than treating them as standalone chat components. It promotes the use of agent-based patterns where AI components perform clearly scoped tasks within multi-step processes.
- Structuring LLM usage as an agent: Agatha behaves like an AI agent—it responds to events, interacts with customers, reasons over input, and assists downstream decisions. The SAS Agentic AI Accelerator supports this pattern by providing repeatable structures for building such agents in SAS Viya.
- Governing LLMs and prompts: The accelerator enables organizations to register LLMs in SAS Viya, manage prompts as versioned assets, and reuse them consistently across workflows. This simplifies experimentation while maintaining governance and traceability.
- Integrating with SAS Intelligent Decisioning: LLM outputs can be directly consumed in SAS Intelligent Decisioning, where they are combined with rules and analytical models. This ensures that final actions remain deterministic, explainable, and auditable, while still benefiting from AI-assisted reasoning.
- Simplifying integration with external AI services: Speech-to-text, text-to-speech, and telephony services can be integrated using standardized orchestration patterns. This allows teams to focus on business logic rather than custom integration code, while keeping all decision control inside SAS Viya.
From hackathon concept to enterprise pattern
This project illustrates what can be achieved with SAS Viya and the SAS Agentic AI Accelerator. It shows how similar ideas can be transformed into scalable, production-ready solutions by providing structure, reuse, and built-in governance.
Together, they demonstrate a clear path forward: LLMs' augment interaction and reasoning; SAS Viya governs decisions and actions, and agent-based architectures to bridge analytics and operations.
Turning AI into action
Agatha is more than a conversational AI prototype. It is an example of how financial institutions can move from fraud detection to fraud resolution by combining LLM-driven interaction with SAS Viya decisioning.
By leveraging frameworks such as the SAS Agentic AI Accelerator, customers in any industry can implement similar solutions in a trusted, explainable, and scalable way—turning innovative concepts into enterprise-ready AI systems.
To explore how this approach can be implemented in practice, learn more about SAS Agentic AI Accelerator and discover how to build, govern, and scale AI agents in enterprise environments.
