Amazon AWS AIP-C01: Exam Guide and Six-Week Study Plan

Building a convincing generative AI demonstration and operating a dependable generative AI application are different engineering problems. The gap becomes apparent when the retrieval source is outdated, the model is unavailable in the preferred Region, an agent repeats a billable action, or a safety control rejects a legitimate customer request. The AWS Certified Generative AI Developer – Professional exam, AIP-C01, tests the decisions that connect these problems rather than treating Amazon Bedrock as a collection of product names.

The AWS Certified Generative AI Developer – Professional (AIP-C01) exam is aimed at developers who can integrate foundation models into working business applications on AWS. This guide sets out its current format, what each official domain expects, a six-week practical preparation sequence, and the kinds of architecture questions worth rehearsing before booking.

AIP-C01 at a glance

The scored and unscored questions are mixed; AWS does not identify which are unscored while you are taking the exam. Multiple-response questions require every correct option, and unanswered questions count as incorrect. There is no penalty for guessing. Because scoring is compensatory, you need to meet the overall passing standard rather than a separate pass mark in every domain. Check the AWS certification website immediately before scheduling, particularly if you need a specific language or testing accommodation.

Who should take it?

AWS describes the target candidate as someone with at least two years of experience developing production-grade applications on AWS or with open-source technologies, some AI/ML or data engineering familiarity, and around a year of hands-on generative AI implementation. That is a profile, not a formal eligibility requirement. AWS does not require another certification before AIP-C01. Its suggested foundations include cloud compute, storage and networking, identity and security, infrastructure as code, observability and cost management.

The boundary matters. AIP-C01 does not primarily assess inventing new model-training algorithms or performing broad data science research. It concentrates on selecting and integrating models, preparing data for inference and retrieval, building agentic applications, securing their connections, evaluating outputs and operating them reliably. If your experience is mostly conceptual, understanding the difference between AIP-C01 and AIF-C01 can clarify whether the Professional path is the right immediate step.

Understand the official exam domains

The weight distribution is unusually useful for planning because the first two domains together make up 57% of scored content. That does not mean the remaining areas are optional: security, operating cost and evaluation are often embedded within an architecture question about model or agent implementation.

Domain 1: Choose the model and earn the right to use the data

The first domain is not simply a test of which foundation models Amazon Bedrock offers. A candidate needs to consider task type, quality benchmarks, supported input modalities, latency, context window, regional availability, data sensitivity and operational cost. A smaller model may be better for a narrow classification task; a more capable model may be justified for complex synthesis. The decision is only defensible if it has measurable acceptance criteria. Our foundation-model selection method provides a practical decision framework.

Retrieval-augmented generation deserves a separate lab because failures frequently begin before inference. Source systems must have owners, document updates need to reach the index, chunks should preserve meaning, embeddings must suit the corpus, and search results must obey the user’s access boundary. Test keyword, semantic and hybrid retrieval on questions that contain acronyms, near-duplicate policies, and ambiguous dates. Reranking can improve relevance but cannot repair missing or stale material. Treat each retrieved document as evidence, not as an instruction authorized to control an agent. For the upstream validation, extraction and media-format decisions, see data quality and multimodal GenAI pipelines for AIP-C01.

Effective prompting and grounding connects these design choices to versioned prompts, permission-aware retrieval and quality regression. For AIP-C01 practice, describe how you would detect that the source is stale and how you would safely refuse or escalate when the evidence is insufficient.

Domain 2: Integrate models with dependable applications

Model APIs, agents and tools need the same production discipline as ordinary web services. AWS Lambda, Amazon API Gateway, Amazon EventBridge and AWS Step Functions can form useful boundaries around an application, but the right pattern depends on task duration, retries, side effects and user expectations. A synchronous model call is sensible for short interactive requests. A lengthy document workflow might need an asynchronous event or orchestration flow with a status record and timeout policy.

Agentic workloads make deterministic control more important, not less. The agent can suggest a tool invocation, but the receiving service should validate parameters, authorize the operation, enforce idempotency and record an auditable result. An instruction telling the model not to issue a refund twice is not a substitute for an idempotency key in the business API. Work through the agents and tool integration guide, then build a small workflow that demonstrates approval, timeouts and safe retries.

Amazon Bedrock can provide managed access to foundation models and related capabilities, but not every application requirement belongs inside Bedrock. For example, business transaction rules should remain in the system of record, while an application can use Bedrock to interpret a request or produce a grounded summary. The Bedrock model-layer architecture article examines where the managed AI layer ends and conventional application engineering begins.

Domain 3: Treat safety as a system design choice

Safety on a production GenAI application involves a chain of trust. User identity, model access, retrieval permissions, prompt templates, tool roles and output handling all need deliberate controls. Amazon Bedrock Guardrails can help filter unwanted content and sensitive information, but guardrails do not replace IAM, validation in downstream services, least-privilege tool design or human authorization for consequential actions.

Prepare cases in which an untrusted document contains instructions to ignore system policy, a support agent tries to retrieve another customer’s data, or a tool requests a secret as part of its parameters. Explain the enforcement point for each case. You should be able to defend both the initial security posture and the evidence retained for investigation. These issues surface in production GenAI security decisions involving concrete application scenarios.

Domains 4 and 5: Measure the production system, not just the answer

Cost optimization can change behavior. Shortening retrieved context lowers token spend, but if it removes the document’s eligibility clause the resulting answer can become unsafe. Switching to a less expensive model may reduce average latency but increase escalations or retries. Test the whole application against business outcomes and quality criteria before treating a lower inference bill as an improvement.

Measure requests and output tokens by use case, model route, cache hit rate, time to first token, tool failures, retrieval success, quality scores, escalations and cost per completed task. A model can be healthy while an agent fails because a downstream role lost permission or a vector index became outdated. Good troubleshooting separates each layer. The GenAI evaluation and optimization process exposes the quality–performance relationship, while troubleshooting production failures focuses on identifying the component that actually broke.

A six-week preparation plan

This is a practical study sequence, not an official AWS completion time. It assumes you have already written and deployed applications. If AWS networking, IAM or Python SDK fundamentals are unfamiliar, extend the preparation period rather than rushing directly to agent frameworks.

Build one small application that can fail in five useful ways

Use a customer-support knowledge assistant with a controlled corpus, one narrow ticket-creation tool and a test identity. Start with a successful grounded question. Then change one input at a time: replace a current policy with an older version; ask for a document the user is not authorized to read; force the tool to time out after completing the action; throttle the model endpoint; and increase retrieval context until costs rise. Capture traces, expected behavior and the real result for each case.

This single project tests more architectural judgment than six unrelated product tutorials. It gives you evidence of what you understood about retrieval, identity, side effects, observability, quality and optimization. The hands-on scenario collection provides additional failure cases to rehearse.

How to answer architecture scenarios under time pressure

Read the constraint before reading product names. Ask whether the requirement is dominated by freshness, permissions, cost, latency, throughput, transactional correctness, quality or recovery. Eliminate options that violate a hard constraint, even if they appear technologically sophisticated. Then compare the remaining options by complexity and operational burden. A design that adds an agent where a validated function would suffice often introduces more failure modes without improving the outcome.

For multiple-response questions, evaluate each option against the entire scenario; some choices may be independently true but inappropriate in combination. Use the official AWS practice assessment to understand exam wording and the skills-measured outline to detect which domain you have neglected. Avoid treating unofficial question counts, promised pass rates or recycled memorization material as substitutes for working knowledge.

Before booking AIP-C01

Verify that the exam page still lists the format, fee, delivery options and languages you need. Confirm test-center or online-proctoring requirements and identity-document rules for your country. AWS offers an exam preparation plan and official practice resources through AWS Skill Builder. AIP-C01 does not require another certification, though AWS suggests foundational experience or credentials as possible preparation.

A good readiness check is whether you can explain not only how a solution uses Amazon Bedrock, but why it is structured that way, what happens when a component fails, who is allowed to trigger a side effect and how you would know whether a lower-cost alternative still produces acceptable business results. Those are the skills the Professional exam is designed to recognize.