Anatomy of the GenAI Lens
Lesson 2 of 5 in The Generative AI Lens: AWS Well-Architected for GenAI.
Open the lens and the first thing you meet is its ID grammar. Every best practice has a name like GENOPS01-BP01, and the name parses mechanically: a pillar prefix (GENOPS for operational excellence, GENSEC for security, GENREL for reliability, GENPERF for performance efficiency, GENCOST for cost optimization, GENSUS for sustainability), a question number within that pillar, and a best-practice number within that question. So GENSEC05-BP01 is the security pillar, fifth question (excessive agency, as it happens), first best practice. Learn the grammar once and you can navigate the whole document — and, more usefully, cite findings unambiguously in tickets, control registers, and crosswalk tables.
Each pillar keeps its classic Well-Architected meaning but acquires a distinctly GenAI set of worries:
| Pillar | Prefix | What it worries about in a GenAI workload |
|---|---|---|
Operational excellence |
| Consistent model output quality and how you verify it; operational health across every layer; traceability and observability; automating lifecycle management; deciding when to customise a model. |
Security |
| Protecting model endpoints; stopping harmful, biased, or ungrounded responses; monitoring and auditing events; prompt security; excessive agency in agentic workflows; data-poisoning risks. |
Reliability |
| Throughput quotas — foundation models can only serve so many inference requests at once; network reliability between components; graceful failure and recovery; prompt version management; distributed availability and compute-task completion. |
Performance efficiency |
| Capturing and improving model performance in production against ground-truth benchmarks; load testing and parameter tuning; compute optimisation; retrieval performance — the vector store as a first-class bottleneck. |
Cost optimization |
| Model selection as a cost decision; understanding GenAI pricing models; engineering prompts for cost; cost-informed vector stores and agent workflows with hard exit conditions. |
Sustainability |
| The computational footprint of training, customisation, hosting, and storage; energy-efficient models and infrastructure; asking whether generative AI is the right tool at all — a search index answers some questions more sustainably than a model. |
The lens organises its guidance along a six-stage generative AI lifecycle: scoping → model selection → customization → development → deployment → continuous improvement. That ordering carries an argument. Scoping — should this be GenAI at all, and what must it do? — comes before any model is touched, and continuous improvement means the review never really ends: new candidate models and fresh customisations send you back through evaluation. If the lifecycle sounds familiar, it should — it is the same shape as the risk-management loops you met in NIST’s AI RMF and ISO/IEC 42001, drawn by engineers instead of policy writers.
Scope matters too, in both directions. The lens covers workloads built on Amazon Bedrock (managed foundation models), Amazon SageMaker AI (customer-managed models), and Amazon Q (packaged business applications). It explicitly does not cover classical predictive ML — that remains the older Machine Learning Lens. If your estate runs both a fraud classifier and a RAG chatbot, you have two lenses to schedule, not one.
One more structural idea runs through the lens: responsibility is split across three roles in the model supply chain. The model producer builds and trains the foundation model. The model provider hosts and serves it — Amazon Bedrock playing this role for models it offers. The model consumer builds applications on top. The lens addresses you mainly as a consumer (and sometimes provider, if you self-host on SageMaker AI), which is why its best practices concern evaluation, guardrails, and access control rather than pre-training data curation — the producer’s problems live upstream, out of your reach but not out of your risk.
You met this idea as the Shared responsibility model matrix in the cloud-guardrails module; the lens gives it a supply-chain vocabulary. The same organisation can hold two roles at once — fine-tune an open-weight model and host it yourself, and you have become provider as well as consumer, inheriting a second role’s worth of best practices.
Key terms: model producer, model provider, model consumer, AI lifecycle, Shared responsibility model
Interactive checkpoint quiz (2 questions) — open this page in a browser to take it.