Announcing the Well-Architected FSI Lens updated for Generative AI and Agentic AI

Today, we’re excited to announce an update to the AWS Well-Architected Financial Services Industry Lens with comprehensive guidance for generative AI and agentic AI workloads. This version is available in the Lens Catalog of the AWS Well-Architected Tool.

What’s New

The updated Financial Services Industry (FSI) Lens introduces new best practices specifically designed for generative AI and agentic AI systems, spanning all six pillars of the AWS Well-Architected Framework. These additions provide financial services organizations with actionable guidance to design, deploy, and operate AI workloads with confidence.

The lens emphasizes treating generative AI and agentic AI as first-class workload components requiring the same rigor in governance, security, performance optimization, and cost management as traditional cloud infrastructure as AI becomes increasingly central for financial services customers.

Introduction

The AWS Well-Architected (WA) Framework has been helping AWS customers and partners improve their cloud architectures since 2015. The AWS WA Framework includes domain-specific lenses, hands-on labs, and the AWS Well-Architected Tool which is available at no cost in the AWS Management Console and provides guidance across six pillars:

Security

Security remains paramount in financial services, and our new guidance addresses the unique challenges of AI systems:

• Secure AI/ML models and protect training data with comprehensive controls including least privilege access, data purification filters for training data integrity, and version-controlled prompt catalogs for secure model deployment
• Monitor AI system outputs for security issues through automated response validation with guardrails, prompt injection detection, and AI-specific incident response procedures
• Implement AI model governance and access controls with comprehensive governance frameworks, model approval workflows, and separation of duties between prompt engineering and security administration
• Leverage AI for threat detection and security automation to enhance anomaly detection, automate malware analysis, and enable AI-driven incident remediation
• Implement fine-grained permission models for agent actions with principle of least privilege
• Define clear security boundaries for agent operations and establish governance for tool access
• Implement safeguards against agent prompt injection and manipulation

Operational Excellence

Financial institutions need robust operational frameworks for AI systems. The updated lens includes:

• Define generative AI model risk management frameworks with comprehensive model inventory, risk tiering, data governance for training datasets, compliance tracking, and guardrails for system behaviors
• Implement human-in-the-loop validation for critical AI processes with mandatory review workflows and comprehensive audit trails
• Establish AI model versioning and rollback strategies using immutable model registries for reliable deployment management
• Add specialized AI system testing including prompt/response testing and adversarial testing for vulnerabilities
• Establish dedicated governance structures for autonomous agents with clear boundaries, permissions, and escalation paths
• Implement specialized monitoring for agent activities and decisions with human oversight thresholds
• Define processes for agent deployment, versioning, and retirement

Reliability
The lens provides comprehensive guidance for ensuring AI system reliability:

• Design resilient AI architectures with appropriate Recovery Time Objectives (RTO) and Recovery Point Objectives (RPO) for AI workloads
• Implement AI model versioning and rollback strategies using immutable model registries
• Add specialized AI system testing and validation including prompt/response testing and adversarial testing for vulnerabilities
• Design resilient agent architectures with graceful degradation and failover mechanisms
• Implement validation mechanisms to ensure consistent agent behavior
• Define procedures for recovering from agent failures or incorrect decisions, and create comprehensive testing frameworks for agent behaviors under various conditions

Performance Efficiency
The lens provides detailed guidance on selecting and optimizing AI infrastructure:

• Select appropriate GPU and accelerated computing with recommendations for instances for model training, real-time inference, and cost-effective transformer inference
• Select and optimize generative AI components covering ground truth datasets for financial use cases, model customization through fine-tuning, and vector store optimization for financial data retrieval
• Define ground truth datasets with financial services-specific scenarios for accurate model evaluation
• Optimize AI model inference performance through inference acceleration techniques including pruning and quantization, real-time optimization for fraud detection and trading, and comprehensive performance monitoring
• Monitor and tune AI system performance with accuracy metrics, business impact tracking, and automated tuning capabilities
• Design efficient workflows for multi-agent systems
• Implement efficient context management for long-running agents and optimize how agents select and use tools
• Design patterns for parallel agent operations when appropriate, and balance thoroughness with response time in agent operations

Cost Optimization
Understanding and controlling AI costs is critical for sustainable adoption. New best practices include:

• Use credits and investment programs for AI/ML proof-of-concepts to reduce initial experimentation costs
• Apply Savings Plans to generative AI inference endpoints and model serving for predictable cost management
• Consider AI-specific pricing trade-offs with model selection based on price-performance ratios, model routing rules and serverless RAG orchestrators
• Track generative AI KPIs including cost per tokens, cache hit percentage, and model tier mix ratios
• Measure AI business impact through cost-to-value metrics such as cost per contract reconciled or cost per query resolved
• Implement strategies to minimize token consumption in agent operations
• Monitor and optimize costs associated with tools that agents invoke
• Design cost-aware agent workflows that minimize unnecessary steps
• Implement appropriate caching for repetitive agent tasks, and define criteria for selecting appropriate agent complexity tiers based on task requirements

Sustainability
The update addresses the environmental impact of AI workloads:

• Select lower carbon regions for AI training workloads to minimize environmental impact
• Monitor token usage and scale down inference endpoints during idle periods for resource efficiency
• Use managed spot training and parameter-efficient fine-tuning (PEFT) techniques to reduce computational requirements
• Benchmark energy-efficient instances for sustainable AI operations
• Develop multi-architecture AI containers for different instance types to optimize resource utilization
• Monitor and optimize computational resources used by autonomous agents
• Implement sustainability-aware task scheduling for non-time-critical agent operations
• Define sustainability KPIs specific to agent operations
• Develop patterns that minimize environmental impact of agent operations, and select energy-efficient infrastructure for agent workloads when possible

Conclusion

The updated AWS Well-Architected Financial Services Industry Lens is available now and we encourage all financial services organizations exploring or deploying generative AI and agentic AI solutions to review this guidance.

For detailed guidance and implementation examples, access the complete FSI Lens documentation through the AWS Well-Architected Tool or the AWS Well-Architected Framework.

Contact your AWS account team to engage a Financial Services Industry specialist if you require additional expert guidance.

Learn more about AWS for Financial Services, customer case studies, and additional resources on our Financial Services website.

Additional resources:

• AWS Well-Architected – Getting started video
• AWS Well-Architected Tool
• AWS Well-Architected Framework
• AWS Well-Architected Labs