1.4.1. The Well-Architected GenAI Lens
💡 First Principle: The six Well-Architected pillars apply to GenAI just as to any AWS workload — but the specific risks in each pillar shift. For GenAI, reliability means handling non-determinism and model API failures; security means preventing prompt injection, not just network intrusion; cost optimization means managing token consumption, not just instance rightsizing.
GenAI-specific risks by pillar:
| Pillar | Traditional Risk | GenAI-Specific Risk |
|---|---|---|
| Operational Excellence | Deployment failures | Prompt regression, model drift, output degradation |
| Security | Unauthorized access | Prompt injection, data exfiltration via LLM, PII leakage |
| Reliability | Service downtime | Model API throttling, hallucination, context overflow |
| Performance Efficiency | High latency, over-provisioning | Token costs, context window waste, cold embedding lookups |
| Cost Optimization | Idle resources | Unnecessary FM invocations, oversized models for simple tasks |
| Sustainability | Energy inefficiency | Wasteful large-model use for simple tasks |
The GenAI Lens best practices hierarchy for AIP-C01 scenarios:
- Use the smallest model capable of the task (Cost + Sustainability)
- Implement semantic caching to avoid redundant FM calls (Cost + Performance)
- Apply Guardrails at both input and output (Security + Reliability)
- Ground responses in retrieved context for factual accuracy (Reliability)
- Log all FM interactions for audit and regression testing (Operational Excellence)
⚠️ Exam Trap: "Most cost-effective" and "best performance" often point to different answers on this exam. When both requirements appear in the same question, look for caching, model tiering, and provisioned throughput as the solutions that satisfy both simultaneously.
Reflection Question: A GenAI application's monthly AWS bill triples after launching. The team used on-demand Bedrock invocations with no caching. What are the three architectural changes, in priority order, that would have the highest cost reduction impact?
