ExamPark
Course menuAIP-C01

Exams / AIP-C01 / Study guide

AWS Certified Generative AI Developer – Professional (AIP-C01) study guide

What the AIP-C01 exam tests, domain by domain, and how to tell the right answer from the nearly right one. Each topic links to fuller study notes and free practice questions.

Last checked against the official exam guide and AWS documentation · How we write our content

About the AIP-C01 exam

The AWS Certified Generative AI Developer – Professional exam checks whether you can put foundation models into real applications on AWS: retrieval-augmented generation (RAG) over company data, agents that call tools, prompts that stay consistent, safety and privacy controls, and the monitoring, cost tuning and evaluation that production needs. AWS aims it at people with 2 or more years of building production applications on AWS or with open-source tools, general AI/ML or data engineering experience, and 1 year of hands-on generative AI work.

It is a developer exam, not a data science one. AWS lists model development and training, advanced ML techniques, and data and feature engineering as out of scope. You need to know how to choose, call, ground, guard, scale and test models, not how to build them.
Questions 75: 65 scored and 10 unscored, which aren’t marked
Time 180 minutes
Question types Multiple choice (one answer from four) and multiple response (two or more from five or more)
Score Scaled from 100 to 1,000; 750 to pass
Scoring model Compensatory: you need to pass overall, not in each domain
Guessing An unanswered question counts as wrong, and there is no penalty for a wrong answer, so answer everything

Check the official certification page for the current price, languages and booking options before you book.

What the exam really tests

Questions describe a team, an application and a set of requirements, then offer four answers that all sound reasonable. The best answer meets every requirement, including the qualifier, using managed features rather than custom code. A few themes run through the whole exam:

  • Ground answers in your data. Knowledge bases, vector stores, chunking, embeddings, metadata filters, hybrid search and reranking.
  • Build agents safely. Tools and MCP, Amazon Bedrock AgentCore, memory, stopping conditions, scoped permissions and human approval.
  • Guard every request. Amazon Bedrock Guardrails, prompt injection defenses, PII protection, grounding checks and audit logging.
  • Spend tokens wisely. Smaller models for simple work, prompt caching, batch inference, service tiers and cross-Region inference.
  • Measure quality. Bedrock evaluations, LLM-as-a-judge, RAG and agent evaluations, quality gates and troubleshooting.

Services that changed after the exam guide was written

The exam guide was published before several AWS changes in 2026, and some services it names now take no new customers. Our questions follow AWS’s current recommendations, because a real project starting today couldn’t use the closed services:

In the exam guide Status since July 30, 2026 Use instead for new work
Amazon Bedrock Agents (now Agents Classic) Closed to new customers Amazon Bedrock AgentCore with a framework such as Strands Agents
Amazon Kendra Closed to new customers Amazon Bedrock Knowledge Bases, including the fully managed knowledge base
Amazon Q Business Closed to new customers Amazon Quick
SageMaker AI Clarify, Model Monitor, Ground Truth and Augmented AI Closed to new customers Amazon Bedrock evaluations, AgentCore Evaluations, CloudWatch, and Step Functions for human review

Amazon Bedrock Prompt Flows is now called Amazon Bedrock Flows. The real exam may still mention the older names, so recognize them, but don’t pick a closed service for a new build when a current one fits.

How to read a question

  1. Read the last sentence first to find the qualifier: least operational overhead, most cost-effective, lowest latency, most secure, fewest code changes.
  2. List the hard constraints: data residency, latency, budget, who may see which documents, how fresh the data must be, existing tools.
  3. Eliminate answers that build custom code where a managed feature exists, send data where it may not go, or solve a different problem (fine-tuning to add fresh facts, for example).
  4. Between the last two, pick the one that meets every requirement with the fewest moving parts.

Pace yourself: 180 minutes for 75 questions is about 2 minutes 24 seconds each. Flag long scenarios and come back.

How the domains add up

Foundation Model Integration, Data Management, and Compliance is the largest domain at 31%, followed by Implementation and Integration at 26%. Together they are more than half the exam, so RAG, prompts and agents deserve most of your time. AI Safety, Security, and Governance is 20%, Operational Efficiency and Optimization 12%, and Testing, Validation, and Troubleshooting 11%. The domains overlap: a troubleshooting question often turns on chunking, and a cost question on model choice.

Domain 1: Foundation Model Integration, Data Management, and Compliance 31%

The largest domain, at 31% of scored content. It covers designing generative AI solutions and proofs of concept, choosing foundation models and switching between them without code changes, staying available through cross-Region inference and fallbacks, fine-tuned models and their lifecycle, validating and preparing text, image, audio and tabular data, vector stores and metadata, keeping a knowledge base in sync with its sources, chunking, embeddings, hybrid search and reranking, query rewriting, exposing retrieval as a tool, and managing, testing and chaining prompts.

Domain 2: Implementation and Integration 26%

At 26%, this domain is about building and connecting the application. It covers agents (Strands Agents and Amazon Bedrock AgentCore), reasoning patterns, safeguards that stop runaway agents, multi-model coordination, human review, tools and MCP servers, choosing between on-demand, Provisioned Throughput and SageMaker AI endpoints, hosting open-weight models on GPUs, model cascading and routing, event-driven integration with enterprise systems, secure access, hybrid and edge deployments, CI/CD and GenAI gateways, synchronous, asynchronous and batch calls, streaming, retries and fallbacks, front ends and developer tools.

The loop an AI agent runs to answer a request A user sends a request to an agent. The agent asks a foundation model to plan the next step. The model can call MCP tools, a knowledge base, or a human approval step for risky actions, and the results go back to the agent, which repeats until it can answer the user. User request answer Agent AgentCore Runtime, Strands Foundation model plans the next step MCP tools e.g. order API Knowledge base retrieval Human approval risky actions results
An agent runs a loop. The model reads the request and picks a tool, the agent runs it and passes back the result, and the model decides again until it can answer. Tools are exposed through MCP servers or action groups. Safeguards keep the loop in check: a maximum number of steps, least-privilege tool permissions, and human approval before actions that change data or spend money.

Domain 3: AI Safety, Security, and Governance 20%

At 20%, this domain covers keeping inputs and outputs safe with Amazon Bedrock Guardrails and layered controls, reducing hallucinations with grounding checks and structured outputs, defending against prompt injection, isolating workloads with VPC endpoints, IAM and encryption, protecting personal data, governance through model cards, lineage and audit logs, continuous monitoring for misuse and drift, and turning responsible AI principles (transparency, fairness, policy compliance) into working controls.

Guardrails checking a generative AI request on the way in and the way out A user prompt passes an input guardrail that checks for prompt attacks, denied topics and personal data, and is either blocked with a safe message or sent to the foundation model with retrieved context. The model output passes an output guardrail that checks grounding and masks personal data, and is either blocked or returned as the response. User prompt Guardrail: input prompt attacks, topics, PII Blocked safe message Foundation model with retrieved context Guardrail: output grounding, PII masking Blocked safe message Response
Guardrails check both directions. Input checks catch prompt attacks, denied topics and personal data before the model sees them. Output checks mask PII, apply content filters, and use contextual grounding to block answers the retrieved sources don’t support. A blocked request gets a fixed safe message instead. Guardrails are one layer: IAM, private networking and logging still apply.

Domain 4: Operational Efficiency and Optimization for GenAI Applications 12%

At 12%, this domain is about running generative AI applications efficiently. It covers token efficiency, choosing cheaper models and service tiers, throughput, quotas and capacity, prompt caching and semantic caching, latency, inference parameters such as temperature and top-p, monitoring with CloudWatch and model invocation logs, agent and tool observability, and keeping vector stores fast and healthy.

Domain 5: Testing, Validation, and Troubleshooting 11%

At 11%, this domain covers measuring quality and fixing problems. It covers model evaluation with Amazon Bedrock evaluations and LLM-as-a-judge, RAG and retrieval evaluation, agent evaluation with AgentCore Evaluations, user feedback and reporting, quality gates for model and prompt updates, and troubleshooting context-window limits, API errors, inconsistent prompts and poor retrieval.

A 7-week study plan

This plan assumes about an hour on weekdays and a longer, hands-on session at the weekend. Time follows the domain weights: RAG and data (31%) gets two weeks, implementation and agents (26%) gets two, and safety (20%) gets a full week. If you already ship generative AI features on AWS, five weeks may be enough.

Week Focus Do this
1 Baseline and models Take the free practice exam in timed mode. Then solution design and proofs of concept, choosing foundation models, the Converse API, switching models with AppConfig and inference profiles, cross-Region inference and fallbacks, fine-tuning, LoRA adapters, Custom Model Import and distillation.
2 Data and retrieval Data validation with Glue Data Quality, multimodal processing with Bedrock Data Automation, knowledge bases and vector stores (OpenSearch, Aurora pgvector, S3 Vectors, managed knowledge bases), metadata, sync, chunking, embeddings, hybrid search, reranking, query rewriting, retrieval as a tool.
3 Prompts Prompt Management, versions and approval, conversation memory, prompt techniques (few-shot, chain-of-thought, structured output), prompt testing, Amazon Bedrock Flows and Step Functions chains.
4 Agents and tools Strands Agents, AgentCore Runtime, Gateway, Memory, Identity and Policy, MCP servers on Lambda and ECS, reasoning patterns, stopping conditions, human approval with Step Functions, multi-model routing.
5 Integration and deployment On-demand, batch, Provisioned Throughput and service tiers, SageMaker AI endpoints for open-weight models, synchronous and asynchronous patterns, streaming, throttling and retries, event-driven integration, secure access, hybrid and edge, CI/CD and GenAI gateways.
6 Safety and governance Guardrails (content filters, prompt attacks, denied topics, PII, contextual grounding, Automated Reasoning), defense in depth, VPC endpoints and encryption, Macie and Comprehend PII, invocation logging and CloudTrail, model cards, transparency and fairness.
7 Operations, testing and practice Token efficiency, prompt and semantic caching, latency, inference parameters, CloudWatch monitoring and AgentCore Observability, Bedrock evaluations, RAG and agent evaluations, quality gates, troubleshooting. Then full timed practice exams; review every wrong answer.

How to use practice questions well

  • Practice the length. Do full 75-question sets in 180 minutes.
  • Find the qualifier. Many wrong answers work but cost more, add latency, need more code or miss one constraint.
  • Read the wrong options. They are usually a real feature solving a neighboring problem, or the right feature set up the wrong way.
  • Don’t memorize questions. The real exam won’t reuse them.
  • Aim for 80% or more across a few full timed practice exams before you book.

Get hands-on

In a personal account with a budget alert set: create a knowledge base from a few documents in S3 and query it with citations, attach a guardrail with a denied topic and a PII filter, call a model through the Converse API with streaming and tool use, turn on model invocation logging, run a small LLM-as-a-judge evaluation, and deploy a simple Strands agent to AgentCore Runtime. Delete everything afterward.

Frequently asked questions

Is this the same as the AI Practitioner exam?

No. AWS Certified AI Practitioner (AIF-C01) is a foundational exam about concepts and use cases. Generative AI Developer – Professional is a hands-on build exam with long scenarios about architecture, code paths, safety controls, cost and troubleshooting.

Do I need to know machine learning?

Only at the level of using models. Training and feature engineering are out of scope. You do need to know when fine-tuning helps, how parameter-efficient methods such as LoRA work at a high level, and how to deploy, version and retire a customized model.

Which services are no longer right answers?

From July 30, 2026, Amazon Bedrock Agents (Agents Classic), Amazon Kendra, Amazon Q Business, and SageMaker AI Clarify, Model Monitor, Ground Truth and Augmented AI are closed to new customers. AWS App Runner is also in maintenance, and the X-Ray SDKs are in maintenance in favor of OpenTelemetry. Our questions use the current replacements: AgentCore, Bedrock Knowledge Bases, Amazon Quick, Bedrock evaluations and CloudWatch.

Do I need the associate certifications first?

No, AWS has no prerequisites. Developer – Associate or Machine Learning Engineer – Associate knowledge helps, because this exam assumes you know Lambda, API Gateway, Step Functions, IAM and CloudWatch well.

Are the practice questions the real exam questions?

No. Every question on this site is original, written from the public exam guide and checked against AWS documentation. Using real exam questions breaks the AWS Certification Program Agreement.

Test yourself

The free AIP-C01 practice exam has 75 questions with an explanation for every option. No sign-up needed.

Take the free practice examReview questions untimed

Want to be sure you're ready? The 5 full-length practice sets add 375 new questions, none of them in the free exam. One payment of $14.99, no subscription. See the practice sets