Exams / AIP-C01 / Study guide / GenAI solution design, proofs of concept and the Well-Architected Generative AI Lens
GenAI solution design, proofs of concept and the Well-Architected Generative AI Lens: AIP-C01 study notes and practice questions
What to know about GenAI solution design, proofs of concept and the Well-Architected Generative AI Lens for the AIP-C01 exam, from the Foundation Model Integration, Data Management, and Compliance domain (31% of the exam guide), then 2 free questions to test it. Part of the AIP-C01 study guide.
AIP-C01 gives a business request with constraints such as data sensitivity, how fast facts change, latency and budget. Choose the simplest pattern that meets them, prove it with a bounded experiment, and name the reference architecture.
Choose the pattern from the constraint
Start with the least complex approach and add parts only when measurements show a gap.
Prompt only: the task is formatting, classification or general reasoning, and the answer does not depend on private data.
RAG: answers must reflect private or changing documents and should cite sources. New facts arrive by re-syncing data, not by retraining.
Fine-tuning or customization: the need is consistent behavior, tone or a strict output format that prompting does not hold. Stable knowledge only.
Agent: the system must choose among tools and take several actions. Agents add cost, latency and risk, so use one only when a fixed workflow is not enough.
Scope a proof of concept
A proof of concept is a bounded experiment. Use a test set drawn from real traffic, set go or no-go thresholds before you start, and measure three things: answer quality (human review or LLM-as-a-judge scoring with Amazon Bedrock evaluations), p95 latency, and cost per request. Build it in the Region and with the data classification you plan to use in production.
Reference architecture
Typical building blocks are an API layer (API Gateway or a Lambda function URL), an orchestrator, Amazon Bedrock with Guardrails, a knowledge base for RAG, and CloudWatch for monitoring. The Well-Architected Generative AI Lens gives a shared review checklist, for teams building similar systems.
Deployment pattern
Use case shape
Pattern
Interactive request, short answer
Synchronous API, or WebSocket for streaming
Large document set, no user waiting
Batch inference from S3
Multi-step task that calls tools
Agent on AgentCore Runtime with tools through Gateway
Triggered by a file or event
EventBridge or S3 event to Lambda or Step Functions
Build or managed
Keep managed Bedrock unless a stated requirement, such as a model Bedrock does not offer or GPU control, forces self-hosting.
How to choose
Scenario wording
Choose
“Answers must cite current internal policy documents”
RAG with a Bedrock knowledge base
“Output format must match exactly, and facts are stable”
Fine-tuning or structured output
“Prove quality, latency and cost before the full build”
Bounded proof of concept with thresholds
“Several teams need one shared design review”
Well-Architected Generative AI Lens
“Data must stay in the EU”
EU Region or EU geographic inference profile
Common traps
Fine-tuning for facts that change weekly, when RAG picks up new documents without training.
Building an agent for a fixed two-step workflow that a Lambda chain handles.
Choosing a model from a public leaderboard without measuring it on your own data.
Designing in one Region and moving data later, which breaks residency and adds transfer cost.
Self-hosting by default when no requirement calls for GPUs or an unsupported model.
A regional bank's compliance team is building an assistant that answers staff questions about 4,000 internal policy documents.
The policies change every week, and each answer must cite the clause it came from. The team also needs every answer to follow one fixed memo layout that the legal reviewers have approved.
A policy change must be reflected in answers within one business day, and the team does not want to retrain a model every week. The team has no staff to run or tune a vector database.
Which architecture meets these requirements with the LEAST retraining effort?
Choose one answer.
AA. Fine-tune a foundation model every week on the policy documents, with the memo layout included in the training examples, and point the assistant at the newest model version.
BB. Place the full text of all 4,000 documents into every prompt, with a fixed system prompt that describes the memo layout, and send each request to a single foundation model.
CC. Create an Amazon Bedrock Managed Knowledge Base over the policy bucket, run an incremental sync after each weekly update, and keep the memo layout in a Prompt Management template.
DD. Fine-tune a model once on the memo layout, then rely on the model's built-in knowledge to answer questions about the bank's current policies without any retrieval step.
Show the answer and explanation
Answer: C
A. Wrong: Weekly fine-tuning means a training job on every policy change, which is the retraining the team wants to avoid. Fine-tuning also does not give reliable clause-level citations, because the model does not store the source passage it learned from.
B. Wrong: This keeps the facts current because the prompt always holds the latest text. However, the documents far exceed a typical context window and would make every request expensive and slow, so it misses the cost and latency needs.
C. Right: Retrieval reads the latest policy text after each sync, so no model is retrained, and the knowledge base returns source passages for citations. Bedrock manages the vector store, so no database staff are needed, and the prompt template enforces the approved memo format. Learn more: Build a managed knowledge base (opens in a new tab)
D. Wrong: A foundation model has no knowledge of the bank's internal policies, so its answers would be unsupported and could not cite clauses. Changes to policies would also be missed entirely until the model was retrained.
Explanation
Policies change weekly and answers need citations, so retrieval over the documents with incremental sync fits better than retraining. A prompt template handles the fixed memo layout, which is a format need rather than a knowledge need. Full-text prompts are too large and costly, and a model-only approach has no access to the bank's policies.
A managed knowledge base over the policy bucket, synced incrementally after each weekly update, keeps answers current.
A hotel group collects about 30,000 guest reviews across its properties each day. Each night, the group wants a draft reply to every review, written in its own tone, so that front-desk staff can approve and post them in the morning.
No guest is waiting for a reply while it is written, and the group wants the lowest possible cost per reply.
The reviews arrive in a shared storage bucket by early evening, and the drafts must be ready before the first shift starts.
Which solution meets these requirements?
Choose one answer.
AA. Use an API Gateway endpoint backed by a Lambda function that calls the model synchronously for each review as it arrives, so every draft is written the moment its review is received.
BB. Write the day's reviews as JSONL files to Amazon S3, and have EventBridge Scheduler start an Amazon Bedrock batch inference job each evening that writes the drafts to S3.
CC. Deploy an AgentCore Runtime agent that a staff member starts each morning, and have it write one reply at a time as staff open each review.
DD. Host an open-weight model on a SageMaker AI real-time endpoint with instances that run around the clock, and send each review to the endpoint as a synchronous request during the night.
Show the answer and explanation
Answer: B
A. Wrong: Real-time calls are priced at the on-demand rate and risk throttling when thousands of reviews arrive together. No guest is waiting, so the real-time design costs more than the nightly need requires.
C. Wrong: Writing drafts one at a time during the morning shift is slow and puts the work on staff at the time they need to approve replies. The group wants all drafts ready before the shift starts.
D. Wrong: The endpoint would run and bill for instances all day, although it is only used for one night's batch. That idle cost misses the lowest-cost requirement, even though the drafts would be ready on time.
Explanation
Nightly reviews with no waiting guest suit batch inference, which is priced below on-demand calls and can be started on a schedule with the output in S3 before the morning shift. Real-time calls, a staff-driven agent and a round-the-clock endpoint each cost more or run at the wrong time.
Want more on GenAI solution design, proofs of concept and the Well-Architected Generative AI Lens? The 5 full-length practice sets have 5 more GenAI solution design, proofs of concept and the Well-Architected Generative AI Lens questions, 375 in all, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets →
Study GenAI solution design, proofs of concept and the Well-Architected Generative AI Lens