AWS AIP-C01 Generative AI Developer Professional guide for RAG, vector stores, prompts, agents, Bedrock integration, safety, governance, optimization, evaluation, and troubleshooting.
This AWS Certified Generative AI Developer - Professional guide is for builders who need to take foundation models into production AWS applications. AIP-C01 is not a prompt-vocabulary exam. It tests whether you can connect FMs to retrieval, agents, APIs, enterprise systems, governance controls, observability, cost controls, and evaluation workflows without losing data boundaries or operational discipline.
Use this root page as the map. The chapter pages teach the production decisions: when RAG beats fine-tuning, when an agent is justified, how tool permissions should be constrained, where guardrails and evaluation fit, and how to debug a GenAI system when output quality, latency, safety, or cost regresses.
| Item | Guide value |
|---|---|
| Vendor | AWS |
| Exam or credential | AWS Certified Generative AI Developer - Professional |
| Code or shorthand | AIP-C01 |
| Study level | Professional GenAI developer |
| Official role target | GenAI developers who integrate foundation models into applications and business workflows. |
| Official experience cue | AWS describes the target candidate as having 2 or more years building production-grade applications, general AI/ML or data engineering experience, and 1 year of hands-on GenAI implementation experience. |
| Guide shape | Start-here page, study plan, cheat sheet, sample questions, FAQ, resources, and glossary. |
As of May 17, 2026, the current AWS AIP-C01 exam guide lists five scored domains:
| Official domain | Weight | What this guide emphasizes |
|---|---|---|
| 1. Foundation Model Integration, Data Management, and Compliance | 31% | RAG, vector stores, knowledge bases, prompt design, model fit, data boundaries, embeddings, and compliance-aware retrieval. |
| 2. Implementation and Integration | 26% | Agents, tool use, enterprise integration, model deployment, FM APIs, workflows, event patterns, and application architecture. |
| 3. AI Safety, Security, and Governance | 20% | Guardrails, input/output controls, privacy, IAM, responsible AI, governance evidence, and policy-aligned behavior. |
| 4. Operational Efficiency and Optimization for GenAI Applications | 12% | Token efficiency, latency, throughput, caching, model tiering, monitoring, quality drift, and cost/performance tradeoffs. |
| 5. Testing, Validation, and Troubleshooting | 11% | Evaluation datasets, release readiness, RAG debugging, agent/tool troubleshooting, model settings, logs, metrics, and rollback signals. |
AWS also says the exam validates vector-store and RAG architectures, FM application integration, prompt engineering and management, agentic AI, cost and performance optimization, security and Responsible AI practices, monitoring and troubleshooting, and FM quality evaluation. Treat those as the guide boundary.
| Lane | What to master | Common weak answer |
|---|---|---|
| Architecture fit | Choose direct inference, RAG, knowledge bases, agents, workflow orchestration, prompt management, model customization, or fine-tuning. | Using a larger or customized model when grounding, retrieval, prompt structure, or deterministic workflow would solve the problem. |
| Retrieval and data boundaries | Manage source data, chunks, embeddings, vector stores, metadata filters, retention, encryption, access, and audit trails. | Treating the vector store as a harmless cache instead of part of the sensitive data boundary. |
| Agent and integration design | Connect Bedrock, tools, APIs, Step Functions, EventBridge, queues, business systems, and enterprise apps safely. | Letting an agent call tools broadly when a narrower workflow, permission boundary, or human approval is required. |
| Safety and governance | Apply guardrails, IAM, private access, prompt-injection defenses, content filters, privacy controls, Responsible AI practices, and compliance evidence. | Treating safety as one feature instead of layered runtime, data, identity, and governance controls. |
| Operations and evaluation | Measure latency, cost, token use, quality, hallucination risk, retrieval relevance, tool failures, version drift, and rollback readiness. | Tuning prompts randomly before checking data freshness, retrieval quality, model parameters, logs, metrics, and evaluation sets. |
Most AIP-C01 questions can be placed in this loop.
flowchart LR
U["User or business workflow"] --> P["Prompt, context, and policy"]
P --> R["Retrieve or call tools when needed"]
R --> M["Foundation model invocation"]
M --> S["Safety, grounding, and structured output checks"]
S --> A["Application action or response"]
A --> O["Observe quality, cost, latency, safety, and drift"]
O --> E["Evaluate, tune, roll back, or govern"]
E --> P
If the prompt asks for current facts, retrieval quality matters before model size. If the prompt asks for action, tool permissions and approval paths matter before agent sophistication. If the prompt asks for production readiness, evaluation and observability matter before launch.
| Scenario pressure | Strong AIP-C01 reasoning |
|---|---|
| Answers are plausible but stale | Check RAG data freshness, indexing, chunking, embedding model, metadata filters, retrieval ranking, and citation behavior before changing the FM. |
| The app must take business actions | Use agent or workflow design only with scoped tools, idempotency, state handling, IAM boundaries, approvals, and audit logging. |
| Sensitive documents are used for grounding | Protect the source, chunks, embeddings, vector store, prompt context, logs, outputs, retention, and cross-tenant access path. |
| Latency or cost is too high | Separate model time, retrieval time, orchestration time, token volume, model tier, caching, batching, streaming, and provisioned capacity decisions. |
| Safety incidents appear in output | Layer prompt constraints, guardrails, retrieval filtering, structured output validation, content review, monitoring, and human escalation. |
| A release improves one metric but hurts another | Use evaluation datasets, baseline comparison, online telemetry, version-aware monitoring, and rollback criteria. |
Use this sequence if you are building the guide from scratch:
AIP-C01 answers should be production-grade: secure the data, ground the output, constrain tools, evaluate behavior, monitor cost and quality, and plan failure handling. Before choosing a model or feature, name the failure mode: stale knowledge, unsafe output, overbroad tool access, weak retrieval, excessive latency, token cost, privacy exposure, or unmeasured quality.
| Trap | Better exam posture |
|---|---|
| Bigger model as the default fix | First check grounding, prompt design, retrieval quality, model settings, and evaluation evidence. |
| Agent for every workflow | Use an agent when autonomous tool choice is needed; use deterministic orchestration when the workflow is known. |
| Fine-tuning for current facts | Use RAG or knowledge bases when the issue is fresh or private source knowledge. |
| Guardrails as the only safety layer | Combine guardrails with IAM, data filtering, prompt controls, output validation, monitoring, and review paths. |
| Logging everything for observability | Keep privacy boundaries around prompts, retrieved context, outputs, embeddings, and user data. |
| Debugging by prompt edits only | Check source freshness, chunking, retrieval, model parameters, API errors, tool execution, logs, metrics, and eval results. |
Use the current AWS exam page for live exam details, including name, status, pricing, duration, delivery method, languages, retirement or beta changes, question types, and domain weights where applicable. This guide follows the official AIP-C01 domain model and uses the local chapter pages as the teaching sequence.