
AWS Certified Generative AI Developer – Professional (AIP-C01) is AWS's professional-level certification for engineers who put foundation models into real applications. It tests whether you can build RAG systems on vector stores, write and govern prompts, build agentic AI with tools and MCP, deploy models, add guardrails and privacy controls, control token cost and latency, and evaluate and troubleshoot GenAI apps in production. Model training and feature engineering are out of scope. The exam is about building with models, not training them.
This guide follows the official AWS exam guide: five domains with weights, the AWS services behind each task, a design cheat sheet, an 8-week plan and 15 original scenario questions.
Quick facts (AWS, November 2026)
| Item | Detail |
|---|---|
| Exam code | AIP-C01 |
| Level | Professional |
| Questions | 75 (65 scored + 10 unscored), multiple choice and multiple response |
| Duration | 180 minutes |
| Passing score | 750 on a 100–1,000 scale (compensatory: you pass on the overall score) |
| Price | USD 300 |
| Recommended experience | 2+ years building production applications on AWS or with open-source tools, plus 1 year of hands-on GenAI work |
Read the official AIP-C01 exam guide and the certification page before booking. New to AWS AI? Start with AWS AI Practitioner (AIF-C01).
Domains and weights
| Domain | Weight |
|---|---|
| 1. Foundation Model Integration, Data Management, and Compliance | 31% |
| 2. Implementation and Integration | 26% |
| 3. AI Safety, Security, and Governance | 20% |
| 4. Operational Efficiency and Optimization for GenAI Applications | 12% |
| 5. Testing, Validation, and Troubleshooting | 11% |
Domains 1 and 2 are 57% of the score. If you can design a RAG pipeline end to end and wire an agent to tools safely, you have most of the exam covered.
Domain 1: FM integration, data and compliance (31%)
- Design and proof of concept: architecture to match business needs; PoCs on Amazon Bedrock; the Well-Architected Generative AI Lens.
- Selecting and configuring FMs: benchmarks and limits; switching models without code changes (Lambda, API Gateway, AWS AppConfig); resilience with Step Functions circuit breakers and Bedrock cross-Region inference; fine-tuned models on SageMaker AI with LoRA/adapters and Model Registry.
- Data for FMs: validation (Glue Data Quality, Data Wrangler); multimodal processing (Transcribe, SageMaker Processing); request formatting; enriching inputs with Comprehend.
- Vector stores: Bedrock Knowledge Bases, OpenSearch Service (Neural plugin, sharding, multiple indexes), Aurora pgvector, DynamoDB for metadata; metadata frameworks; keeping indexes fresh with incremental sync.
- Retrieval: chunking (fixed, hierarchical); choosing embeddings (Titan, dimensions); hybrid search and rerankers; query expansion and decomposition; retrieval through function calling and MCP.
- Prompts: Bedrock Prompt Management (parameterised templates, approval), Guardrails, Prompt Flows for chains and branching, CloudTrail for audit, regression tests for prompts.
Domain 2: Implementation and integration (26%)
- Agentic AI: Strands Agents and AWS Agent Squad for multi-agent systems; ReAct with Step Functions; stopping conditions, timeouts and IAM boundaries; human approval steps; MCP servers on Lambda (light tools) or ECS (complex tools).
- Deployment: on-demand vs provisioned throughput vs SageMaker AI endpoints; GPU and memory constraints of LLMs; smaller models and cascading for routine queries.
- Enterprise integration: event-driven designs with EventBridge; identity federation and least-privilege access to models; Outposts and Wavelength for data residency; CI/CD with CodePipeline and a central GenAI gateway.
- FM APIs: sync calls, async with SQS, streaming responses (WebSockets, server-sent events), exponential backoff, rate limits, fallbacks, X-Ray tracing, model routing.
- Developer tools: Amplify UIs, OpenAPI, Bedrock Data Automation, Amazon Q Developer, CloudWatch Logs Insights for prompt and response analysis.
Domain 3: Safety, security and governance (20%)
- Input and output safety: Bedrock Guardrails on prompts and responses; grounding with Knowledge Bases to reduce hallucinations; JSON Schema for structured output; defence in depth (Comprehend pre-filter, guardrail, Lambda post-check); prompt injection and jailbreak detection.
- Data security and privacy: VPC endpoints, IAM, Lake Formation; PII detection with Comprehend and Macie; masking and anonymisation; S3 Lifecycle for retention.
- Governance: SageMaker model cards, Glue lineage and Data Catalog, CloudTrail audit logs, drift and misuse monitoring.
- Responsible AI: source attribution and agent tracing; fairness checks with A/B tests and LLM-as-a-judge; policy-based guardrails.
Domain 4: Cost and performance (12%)
- Cost: token tracking, context pruning, prompt compression, response limits, tiered models by query complexity, batching, semantic caching and prompt caching.
- Performance: streaming, latency-optimised models, parallel calls, temperature and top-p/top-k tuning, index optimisation.
- Monitoring: CloudWatch for token usage, quality and hallucination rates; Bedrock model invocation logs; tool-call and multi-agent tracing; golden datasets to catch hallucinations.
Domain 5: Testing and troubleshooting (11%)
- Evaluation: relevance, factual accuracy, consistency and fluency; Bedrock Model Evaluations; A/B and canary tests; RAG evaluation; agent task-completion metrics; quality gates in deployment.
- Troubleshooting: context-window overflow and truncation; API integration errors; weak prompts; poor retrieval (embedding quality, chunking, drift); prompt-format drift found with schema validation and X-Ray.
Design cheat sheet
| Requirement | Usual answer |
|---|---|
| Managed RAG with the least code | Bedrock Knowledge Bases with a managed vector store |
| Keyword-heavy queries (product codes, IDs) miss results | Hybrid search (keyword + vector), then a reranker |
| Block PII and harmful content in and out | Bedrock Guardrails (plus Comprehend or Macie upstream) |
| Swap the model without redeploying code | Model ID in AppConfig behind Lambda or API Gateway |
| A model isn't available in your Region or hits capacity limits | Bedrock cross-Region inference |
| Steady, predictable high volume | Provisioned throughput |
| Same questions asked repeatedly | Semantic cache and prompt caching |
| Agent must not loop forever or exceed its permissions | Step Functions stopping conditions, timeouts and IAM boundaries |
| Compare two prompts or models on quality | Bedrock Model Evaluations or LLM-as-a-judge with an A/B test |
8-week study plan
| Week | Focus | Lab |
|---|---|---|
| 1 | Bedrock APIs (Converse, streaming), model choice, inference parameters | Streaming chat API behind API Gateway and Lambda |
| 2 | Chunking, embeddings, Knowledge Bases, OpenSearch and pgvector | RAG over your own PDFs with Knowledge Bases |
| 3 | Hybrid search, reranking, query rewriting, metadata filters | Add a reranker and measure the difference |
| 4 | Prompt Management, Prompt Flows, Guardrails | Versioned prompt with a guardrail blocking PII |
| 5 | Agents: Strands, tools, MCP, Step Functions, human approval | Agent that calls two tools via an MCP server on Lambda |
| 6 | Security, privacy and governance: VPC endpoints, Macie, model cards, CloudTrail | Private RAG with no internet path |
| 7 | Cost and latency: caching, provisioned throughput, routing; CloudWatch and invocation logs | Token dashboard and a semantic cache |
| 8 | Evaluations, troubleshooting, official practice questions | Evaluate your RAG app; book the exam |
15 scenario questions with answers
These are original scenarios written for this guide, not exam questions.
1. A company wants a chatbot over 20,000 policy documents with the least custom code. What do you use?
Amazon Bedrock Knowledge Bases connected to the documents in S3, with a managed vector store.
2. Searches for exact part numbers return irrelevant chunks. What improves retrieval?
Hybrid search combining keyword and vector scores, followed by a reranker model.
3. Answers must cite only current documents after nightly updates. What do you configure?
Incremental sync of the knowledge base data source on a schedule or on change events.
4. The team wants to test a new model without changing application code. How?
Keep the model ID in AWS AppConfig and read it in the Lambda function that calls Bedrock.
5. A model is only offered in some Regions and requests are being throttled. What helps?
Bedrock cross-Region inference, with retries using exponential backoff.
6. Customer Aadhaar and phone numbers must never appear in responses. What do you add?
Bedrock Guardrails with sensitive-information filters (mask or block), plus PII detection on stored data.
7. Users try "ignore your instructions and…" attacks. Which control applies?
Guardrails prompt-attack filters with input sanitisation, plus least-privilege tool permissions so a successful injection can do little.
8. An agent sometimes loops calling the same tool. How do you stop it safely?
Stopping conditions and a maximum step count in the orchestration (for example, Step Functions), with timeouts on tool calls.
9. Refunds above ₹50,000 suggested by an agent need a manager's approval. What pattern?
A human-in-the-loop approval step in Step Functions before the action runs.
10. Users wait 12 seconds for long answers. How do you improve perceived latency?
Stream the response with the Bedrock streaming API over WebSockets or server-sent events.
11. Token costs doubled; many questions are near-duplicates. What do you add?
A semantic cache (and prompt caching for long shared prefixes), plus routing simple queries to a smaller model.
12. The workload is steady at high volume all day. Which pricing option?
Provisioned throughput.
13. Answers drift after a prompt change. How do you catch this before release?
Regression tests against a golden dataset with automated evaluation as a deployment quality gate.
14. Long documents give incomplete answers and some requests fail. What do you check first?
Context-window overflow and truncation; fix chunk sizes, retrieve fewer and better chunks, and trim the prompt.
15. Auditors ask which documents an answer came from. What do you build in?
Source attribution from retrieval metadata in every response, with invocation logs and CloudTrail for the audit trail.
After AIP-C01
- Strengthen the foundations with RAG chunking strategies, hybrid search and reranking and RAG evaluation metrics.
- Prepare for interviews with Bedrock Knowledge Bases interview questions and Bedrock AgentCore interview questions.
- Compare with the Databricks Generative AI Engineer Associate and Microsoft AI-103 guides.
- See live AI and cloud jobs.
Cloud Soft Solutions in Ameerpet, Hyderabad teaches Amazon Bedrock, RAG and agentic AI with real-time projects. See the APEX program, the FDE PRO program and all courses.
📞 Call +91 96660 19191 · WhatsApp · Flat No. 513, 5th Floor, Nilagiri Block, Aditya Enclave, Ameerpet, Hyderabad
Frequently Asked Questions
What is the exam code for AWS Generative AI Developer Professional?
AIP-C01.
How many questions are in AIP-C01?
75 questions in 180 minutes: 65 scored and 10 unscored, multiple choice and multiple response.
What is the AIP-C01 passing score?
750 on a scaled score of 100–1,000.
How much does AIP-C01 cost?
USD 300.
Which AIP-C01 domain has the most weight?
Foundation Model Integration, Data Management, and Compliance, at 31%.
Does AIP-C01 test model training?
No. Model development and training, advanced ML techniques and feature engineering are out of scope. It focuses on building applications with foundation models.



