AWS Certified Generative AI Developer - Professional (AIP-C01) - Exam Strategy¶
Cert-specific tactics. General study advice lives in study-strategies.md. This page is what is different about AIP-C01.
Format reminder¶
- 85 questions total, 65 scored and 10 unscored, 205 minutes
- Pass mark 750 / 1000 scaled
- Multiple choice (1 of 4) and multiple response (2+ of 5+)
- Compensatory scoring: you pass on the overall score, not domain by domain
- Multiple-response questions are all-or-nothing. Partial selections score zero
Time management math¶
205 minutes across 85 questions is about 2.4 minutes per question. GenAI scenarios tend to be long, because they have to establish the model, the data, the latency budget, and the compliance constraint before asking anything.
- Short service-recall questions (~20): 60-90 seconds
- Medium architecture scenarios (~45): 2-3 minutes
- Long multi-constraint scenarios (~20): 4-5 minutes
Aim to be at question 30 by minute 70 and question 60 by minute 145, leaving roughly 40 minutes for flagged items and review. Answer every question. There is no guessing penalty, so a flagged guess always beats a blank.
The top traps for this exam¶
-
RAG versus fine-tuning versus prompt engineering. The most common decision on the exam. The deciding constraint is usually in the last sentence. Knowledge that changes frequently or must be cited points to RAG. Consistent tone, format, or a domain vocabulary the base model lacks points to fine-tuning. Behavior achievable with better instructions points to prompt engineering. Cost and latency budgets break ties.
-
"Least operational overhead" versus "lowest cost." These select different answers, exactly as on SAP-C02. Bedrock managed features (Knowledge Bases, Agents, Guardrails) usually win on overhead; a self-managed vector store on OpenSearch or pgvector can win on cost at scale. Read which criterion the question actually states.
-
Guardrails are not the same as IAM. Bedrock Guardrails filter content and block topics. They do not control who can invoke a model. Questions that describe preventing a team from calling a specific model want IAM policies or resource-based policy, not Guardrails.
-
Knowledge Bases does not mean you skip chunking decisions. Managed ingestion still exposes chunking strategy, embedding model choice, and metadata filtering. Questions about poor retrieval quality usually resolve to chunk size, overlap, or a missing metadata filter, not to swapping the foundation model.
-
Provisioned Throughput versus On-Demand. Provisioned Throughput buys committed capacity for predictable high volume and is the answer when the scenario stresses consistent throughput or guaranteed capacity. It is the wrong answer for spiky or low-volume workloads, where On-Demand is cheaper.
-
Agents introduce failure modes that look like model problems. When a scenario describes an agent calling the wrong tool or looping, the fix is usually the action group schema, the tool description, or the orchestration prompt, not a larger model.
-
Evaluation questions want a method, not a vibe. If a question asks how to know whether a change helped, the answer involves a held-out dataset and a defined metric, often LLM-as-a-judge with human spot checks, not "test it in the console."
Domain weighting and where to spend effort¶
| Domain | Weight | Effort note |
|---|---|---|
| 1. Foundation model integration, data management, compliance | 31% | Largest domain. RAG design and data handling dominate |
| 2. Implementation and integration | 26% | Agents, action groups, API integration patterns |
| 3. AI safety, security, and governance | 20% | Guardrails, IAM, PII handling, model access control |
| 4. Operational efficiency and optimization | 12% | Caching, batching, model selection for cost and latency |
| 5. Testing, validation, troubleshooting | 11% | Evaluation methodology, debugging retrieval and agents |
Domains 1 and 2 are 57% of the exam between them. If time is short, make sure RAG architecture and agent implementation are solid before polishing the smaller domains.
Question triage¶
Read the last sentence first on long scenarios. It carries the decision criterion (cost, latency, overhead, compliance) and often makes two of the four options obviously wrong before you have read the setup.
For multiple-response questions, count the required selections and treat each option as an independent true or false claim. Because scoring is all-or-nothing, a careful pass over five options beats a fast guess.
Flag and move on after 90 seconds of no progress. GenAI scenarios reward a second read with a clear head more than most exams, because the constraint you missed is usually a single clause.
The week before¶
- Re-read bedrock-platform-deep-dive.md and rag-architecture-deep-dive.md. Between them they cover the majority of the tested surface.
- Work scenarios.md and write down why each distractor fails, not just which answer is right.
- Skim aws-services-mapping.md as a last-mile check that you can name the service for each capability.
- Do not start new material in the final three days. Consolidate instead.
Exam day¶
Standard logistics are in the exam-day checklist. Two AIP-C01-specific notes:
- The exam is Professional-tier length. Take the offered break if you are at a testing center, and pace hydration accordingly for online proctoring where you cannot leave.
- Expect several questions that feel like they have two right answers. They do. One is merely correct, the other is the AWS-recommended pattern under the stated constraint. Pick the recommended pattern.