Architecture and Model Fit

AIP-C01 does not reward GenAI feature stacking. It rewards choosing the smallest architecture that satisfies the requirement, can be governed, and still works when latency, cost, model availability, data quality, and failure handling become real.

Start with the requirement, not the model

Split the scenario into six decisions before choosing a service pattern:

1. Does the app need current or proprietary facts?
2. Does it need to take actions or only generate content?
3. Are latency and cost tight enough to limit context size or orchestration depth?
4. Does the answer need auditability, citations, or approval checkpoints?
5. Does the workload need model fallback, cross-Region continuity, or graceful degradation?
6. Is the real gap model capability, source-data quality, prompt design, or workflow design?

If you answer those first, most distractors become easier to remove.

Architecture chooser

What Task 1.1 is really testing

The official Domain 1 outline includes requirement analysis, technical proof of concept, and standardized architecture components. In exam terms, that means the best answer usually:

• maps business need to a specific GenAI pattern
• validates feasibility before full production rollout
• uses repeatable AWS components instead of one-off scripts
• includes security, reliability, observability, cost, and governance from the start
• avoids overbuilding an agent, fine-tune, or custom model when retrieval or prompt governance is enough

This is architecture judgment, not just service recall.

When RAG beats customization

RAG is usually stronger when the knowledge changes often, the answer must cite current internal material, or the organization wants lower retraining overhead. Customization becomes stronger when the problem is repeated behavior, domain style, or task execution quality that retrieval alone does not fix.

Do not treat fine-tuning as the universal accuracy answer. If the missing information already exists in changing enterprise documents, retrieval is usually the cleaner first architecture.

Bedrock-specific exam habit

AWS wants you to think in production lanes, not open-ended research lanes. The strongest AIP-C01 answer often uses Amazon Bedrock plus surrounding AWS services for orchestration, monitoring, access control, and deployment consistency instead of inventing a brittle custom control plane.

Model selection is an architecture decision

Selecting an FM is not only “choose the most capable model.” AIP-C01 expects tradeoffs:

If two models both meet quality requirements, the cheaper or faster model may be the stronger answer. If the scenario is regulated or high-risk, auditability and governance can matter more than raw model capability.

Flexible routing and provider switching

Domain 1 explicitly includes flexible architecture patterns for dynamic model selection and provider switching. The exam-friendly pattern is a thin routing layer that keeps model choice configurable instead of hard-coded:

This does not mean every app needs multi-provider complexity. It means high-availability or fast-changing workloads should not require code redeploys just to change a model route, fallback, or prompt version.

Resilient GenAI systems degrade deliberately

If the question mentions service disruption, limited regional availability, strict uptime, or production continuity, think beyond the prompt. Strong answers may include:

• circuit breaker behavior when an FM route fails
• fallback to a smaller or alternate model
• Cross-Region Inference where supported for the selected Bedrock model
• queued or asynchronous handling for noninteractive work
• graceful degradation, such as returning a grounded summary without optional enrichment
• rollback for failed model, prompt, or customization deployments

Weak answers assume the first model endpoint is always available or that retrying forever is an architecture.

Proof-of-concept questions are not throwaways

If the prompt asks what to validate first, do not rush to full deployment architecture. A proof of concept should isolate the risk:

• model capability if you are unsure the FM can perform the task
• retrieval quality if grounding is the main dependency
• latency and cost if usage volume or response-time targets are strict
• integration fit if the app must call internal tools or enterprise systems
• safety behavior if the app has regulated or high-risk output

The POC should have success criteria. “Build a quick demo” is weak. “Use Amazon Bedrock to compare candidate FMs against a fixed evaluation set with latency, cost, and safety thresholds” is much stronger.

Customization has lifecycle cost

Customization can be valid, but the exam expects deployment discipline:

If an answer says “fine-tune the model” but says nothing about evaluation, deployment, rollback, or lifecycle management, it is usually incomplete for this exam.

Common traps

• choosing fine-tuning just because the prompt says “improve accuracy”
• using an agent when a simple retrieval-plus-prompt pipeline is enough
• treating higher parameter count as a complete answer
• ignoring fallback and provider-switch patterns for availability or continuity
• shipping a POC result without production evaluation, monitoring, or rollback criteria

Decision rule

If the question is really about facts, start with retrieval. If it is really about actions, start with agent workflow and tool control. If it is really about repeatable domain behavior, consider customization only after prompt and retrieval fit are tested. If it is still unclear, the safest first step is a Bedrock-based POC with measurable evaluation criteria.

Quiz