What AIP-C01 Practice Questions Actually Look Like
The AWS Certified Generative AI Developer – Professional (AIP-C01) exam is built around scenario-based questions, not definition recall. A typical question describes a real situation — a company building a generative AI feature on AWS — and asks you to choose the best architecture, the most cost-effective option, or the correct way to secure or troubleshoot it. You'll see multiple-choice (one correct answer) and multiple-response (select all that apply) formats across 65 scored questions.
That means the best practice questions don't ask "What is Amazon Bedrock?" They ask "Given these constraints, which combination of services should you use, and why?" This guide shows you realistic sample questions in that style for each exam domain, explains the reasoning, and points you to legitimate practice resources. For the full exam breakdown, start with our AWS Certified Generative AI Developer – Professional study guide and the AIP-C01 exam guide covering all five domains.
A note on integrity: The sample questions below are original, written to mirror the style and difficulty of the exam. They are not real exam questions. Avoid "brain dump" sites that publish actual exam content — using them violates the AWS Certification Agreement and can get your certification revoked.
Where the Questions Come From: The Five Domains
Every AIP-C01 question maps to one of five weighted domains. Knowing the weighting tells you where to expect the most questions:
| Domain | Weight | ~Questions (of 65 scored) |
|---|---|---|
| 1. Foundation Model Integration, Data Management & Compliance | 31% | ~20 |
| 2. Implementation and Integration | 26% | ~17 |
| 3. AI Safety, Security, and Governance | 20% | ~13 |
| 4. Operational Efficiency and Optimization | 12% | ~8 |
| 5. Testing, Validation, and Troubleshooting | 11% | ~7 |
The first two domains are more than half the exam, so most of your practice questions should live there. Let's walk through a realistic sample for each.
Domain 1 Sample: Foundation Models, Data & Compliance
Question. A financial services company wants to build an internal assistant that answers employee questions using thousands of internal policy PDFs. Responses must stay grounded in the company's own documents, and the documents must not be used to train any foundation model. Which approach best meets these requirements?
- A. Fine-tune a foundation model on the policy PDFs and deploy it via a SageMaker endpoint
- B. Implement Retrieval Augmented Generation (RAG) using Amazon Bedrock Knowledge Bases with the PDFs in Amazon S3
- C. Paste the most relevant policy text into the system prompt for every request
- D. Train a custom model from scratch using Amazon SageMaker
Answer: B. RAG with Bedrock Knowledge Bases keeps responses grounded in the company's documents at inference time, and Bedrock does not use your data to train base models — meeting both requirements. Fine-tuning (A) bakes data into the model and is overkill for a document-grounded Q&A use case. Stuffing the prompt (C) doesn't scale to thousands of documents. Training from scratch (D) is enormously expensive and unnecessary.
The RAG-vs-Fine-Tuning Decision
A huge share of Domain 1 questions hinge on choosing between RAG, fine-tuning, and prompt engineering. The rule of thumb the exam rewards: use RAG when answers must come from specific, changing documents; use fine-tuning to change style/behavior or teach a narrow task; use prompt engineering first because it's the cheapest to iterate on.
Domain 2 Sample: Implementation and Integration
Question. A developer is building a customer-support agent on Amazon Bedrock that must look up order status from an internal REST API and issue refunds. Which combination of components correctly implements this with Bedrock Agents? (Select TWO.)
- A. An action group backed by an AWS Lambda function that calls the order API
- B. A direct database connection embedded in the foundation model's prompt
- C. An OpenAPI schema describing the available actions for the agent
- D. A Step Functions state machine replacing the agent entirely
- E. Hardcoded refund logic inside the model's system prompt
Answer: A and C. Bedrock Agents use action groups that map to a Lambda function (A) for executing tasks like calling the order API or issuing refunds, and an OpenAPI/function schema (C) that tells the agent what actions exist and what parameters they take. Models can't hold live database connections (B) or execute real refund logic from a prompt (E), and Step Functions (D) orchestrates workflows but doesn't replace the agent's reasoning loop.
Domain 3 Sample: AI Safety, Security & Governance
Question. A team is launching a public-facing chatbot on Bedrock and must block responses containing hate speech, prevent the model from discussing competitors, and redact any personally identifiable information (PII) from user inputs before processing. Which AWS capability most directly addresses all three requirements?
- A. IAM resource policies on the Bedrock model
- B. Amazon Bedrock Guardrails
- C. A VPC endpoint for Bedrock
- D. Amazon CloudWatch alarms
Answer: B. Amazon Bedrock Guardrails provide content filters (hate speech), denied-topic configuration (competitors), and sensitive-information filters including PII redaction — covering all three needs in one feature. IAM (A) controls who can call the model, not what it says. VPC endpoints (C) handle network privacy. CloudWatch (D) is for monitoring, not content safety.
Domain 4 Sample: Operational Efficiency & Optimization
Question. A production application makes millions of Bedrock calls per day with highly repetitive prompts, and the team needs to reduce both latency and cost without changing the model's behavior. Which optimization should they implement first?
- A. Switch to the largest available model for better efficiency
- B. Implement prompt caching for the repeated prompt prefixes
- C. Move all traffic to on-demand pricing
- D. Disable Guardrails to save processing time
Answer: B. Caching repeated prompt prefixes avoids reprocessing identical context, cutting both latency and cost while keeping behavior identical. Switching to a larger model (A) usually increases cost and latency. On-demand (C) is appropriate for spiky traffic but high-volume steady traffic often benefits from provisioned throughput. Disabling Guardrails (D) trades away safety for marginal savings — never the right answer on this exam.
Cost-Optimization Questions Have a Pattern
When a question asks for the "most cost-effective" option, the answer is rarely "use a bigger model" or "remove a safety control." Look for choices that reduce redundant work (caching), right-size the model to the task, or match the pricing model (provisioned vs. on-demand) to the traffic pattern.
Domain 5 Sample: Testing, Validation & Troubleshooting
Question. After deploying a RAG application, users report that the model frequently returns "I don't know" even for questions clearly answered in the source documents. The retrieval step returns chunks, but they're often irrelevant. What is the most likely fix?
- A. Increase the model's maximum token limit
- B. Tune the chunking strategy and embedding configuration in the knowledge base
- C. Switch to a smaller foundation model
- D. Add more Guardrails
Answer: B. Irrelevant retrieved chunks point to a retrieval-quality problem — usually chunk size/overlap or the embedding model — so tuning the knowledge base's chunking and embeddings is the direct fix. Token limits (A) affect length, not relevance. A smaller model (C) won't improve retrieval. Guardrails (D) control safety, not retrieval accuracy.
How to Practice the Right Way
Working through realistic questions like these is only useful if you do it deliberately:
- Always read the explanation. The goal isn't to memorize that "B is correct" — it's to internalize why RAG beats fine-tuning here, or why Guardrails is the safety answer. The exam reshuffles the same reasoning into new scenarios.
- Track your weak domains. If you keep missing Domain 3 questions, that's where to study — not the domains you already ace.
- Simulate exam conditions. Do full-length, timed sets. The professional-level time pressure is real, and pacing is a skill.
- Aim for understanding, not a score. Hitting 75–80% on practice exams is a good readiness signal, but only if you understand your correct answers too. See our guide to how the AIP-C01 passing score works for what to target.
Where to Find Legitimate Practice Questions
Stick to sources that teach reasoning and respect the certification agreement:
- AWS Skill Builder — Official Practice Question Set. AWS's own exam-aligned questions. This is the single most authoritative source and worth doing first.
- Reputable third-party practice exams with explanations. Providers like Tutorials Dojo are popular because every question includes a detailed explanation and reference links — that's what makes them useful for learning, not just testing.
- Hands-on practice. Nothing beats building. Spin up a Bedrock Knowledge Base, wire up an agent with a Lambda action group, and configure Guardrails yourself. The scenario questions are far easier when you've actually done the thing they describe.
Hands-On Beats Memorization
The AIP-C01 is a developer exam — it assumes you've built real generative AI applications. If you've personally implemented a RAG pipeline and debugged bad retrievals, the Domain 5 question above answers itself. Build first, then drill questions to find your gaps.
Next Steps
Before you book the exam, make sure you meet the recommended prerequisites, understand how scoring works, and have budgeted for the exam cost. For the complete preparation roadmap — study timeline, services to master, and resources — see the AWS Certified Generative AI Developer – Professional study guide.
Ready to build the hands-on skills these questions test? Explore our AI & Machine Learning courses for project-based training in Bedrock, RAG, and production generative AI on AWS.
Want to practice this hands-on?
CloudaQube generates complete labs from a simple description. Try it free.
Get Started Free