Hands-on AWS knowledge base for developers Managed Ops
Talk to Us

AIP-C01 Exam Guide and Study Plan: 4 Domains, 6 Weeks

Export PDF Print-friendly layout with a copyright watermark; choose Save as PDF in the print dialog
AIP-C01 Exam Guide and Study Plan: 4 Domains, 6 Weeks

TL;DR

  • AIP-C01 is a Professional exam: 65 questions in 130 minutes, scaled score 750 to pass, and Domain 1 (Foundation Model Integration, Data Management, and Compliance) is 31% of the score — roughly 20 of those questions.
  • Plan 40–60 hours across six weeks, front-loading Domain 1 and Domain 4, which together are 54% of the exam.
  • Do every lab in one region (us-east-1 for most of mine), enable model access before you touch code, and get comfortable with inference profiles — half the questions I missed in practice runs were configuration questions wearing a scenario hat.

The four domain weights, and who this exam is written for

The public exam guide splits AIP-C01 into four domains. Read the weights before you buy a course, because they don't match the marketing.

# Domain Weight ≈ questions of 65
1 Foundation Model Integration, Data Management, and Compliance 31% ~20
2 Implementation and Integration 26% ~17
3 Operational Efficiency and Optimization for GenAI Applications 20% ~13
4 Security, Compliance, and Governance for GenAI Applications 23% ~15

Domain 1 is not "prompt engineering." It covers model selection trade-offs (on-demand vs provisioned throughput, foundation model vs self-hosted endpoint), data pipelines for RAG, chunking and embedding strategy, and where your data is allowed to live. That framing matters: a 7B model self-hosted on a g5.xlarge and an on-demand Bedrock call sit in completely different cost and latency universes, and the exam expects you to pick correctly for a stated constraint.

This cert is written for people who ship the feature, not for people who sell it. You should already be comfortable with IAM policies, Lambda, S3 event flows and CloudWatch. If you've never written a resource-based policy or debugged a 403 from an API Gateway authorizer, fix that first — Domain 2 and Domain 4 will eat you alive otherwise.

Registration: what it costs, what it looks like, and which exam to take first

AIP-C01 lists at $300 USD at the Professional tier, delivered through Pearson VUE either at a test centre or online-proctored. You get a scaled score from 100–1000, 750 to pass, and the score report lands in your AWS Certification account after the exam. Question formats are multiple choice and multiple response; AWS has also been shipping ordering and matching items in the newer AI exams, so don't be thrown if one appears.

Exam Level Questions / time Pass score List price (USD) Take it if
AIF-C01 AI Practitioner Foundational 65 / 90 min 720 $100 You're new to AWS AI services or non-building
MLA-C01 ML Engineer Associate Associate 65 / 130 min 720 $150 You own SageMaker training pipelines
AIP-C01 GenAI Developer Professional Professional 65 / 130 min 750 $300 You build Bedrock-backed features in production

If you're deciding order: AIF-C01 first if you've never sat an AWS AI exam, then straight to AIP-C01. Skipping to AIP-C01 without the associate-tier vocabulary is doable but you'll spend two weeks learning service names instead of architecture. The official practice exam costs extra and is worth every cent — same question style, same trick phrasing.

When NOT to book AIP-C01 yet

Hold off if any of these is true. You've never enabled model access in the Bedrock console and hit the "model access" wall. You can't explain the difference between a VPC endpoint and a PrivateLink service. You haven't debugged a throttling error from an inference API and had to request a quota increase. Booking an exam date to force motivation works for some people; in my experience it mostly produces a $300 retake, because Professional-tier questions are scenario chains, not definitions. Two weeks of labs first is cheaper than a reschedule.

A six-week study plan that front-loads the 31% domain

Fifty hours, 8–10 per week. Every week ends with something running, not something read.

Week Focus Hours Deliverable
1 Domain 1: model selection, embeddings, chunking, residency 8 Knowledge Base over ~200 PDFs in S3
2 Domain 1: PII handling, model evaluation, Guardrails 8 Evaluation job + guardrail with a denied topic
3 Domain 2: Agents, action groups, async flows, Step Functions 10 Agent calling a Lambda action group
4 Domain 3: latency and cost — streaming, caching, throughput 8 Token-cost and p95 latency notes per model
5 Domain 4: IAM, KMS, VPC endpoints, CloudTrail, SCPs 8 Deny-by-default SCP test that actually fails a call
6 Practice exam, weak-area re-read, timed drills 8 Two 65-question runs under 130 minutes

Start week 1 by checking what your account can actually invoke. Model access is per account, per region, and this one command has saved me from a wasted afternoon more than once:

aws bedrock list-foundation-models 
  --region us-east-1 
  --query "modelSummaries[?modelLifecycle.status=='ACTIVE'].[modelId,providerName]" 
  --output table


Two more study habits that paid off. Re-read the four domain descriptions in the exam guide until you can predict which domain a practice question belongs to — that alone tells you which constraint matters. And keep a running notes file of every error message you hit; the exam is full of scenarios that read like error messages you've already debugged.

## The Bedrock lab that failed at 1:40am, and what it taught me

Week 3 of my prep, 01:40 UTC, a Step Functions workflow calling a Lambda that invoked a Claude model in us-east-1. The Lambda had been working for two days. Then it started returning this:


An error occurred (ValidationException) when calling the InvokeModel operation:
Invocation of model ID anthropic.claude-3-5-sonnet-20240620-v1:0 with on-demand
throughput isn't supported. Retry your request with the ID or ARN of an inference
profile that contains this model.


My first configuration was wrong in two places. The model ID was hard-coded without a cross-region inference profile prefix, and the IAM policy granted `bedrock:InvokeModel` only on the bare foundation-model ARN. I'd copied both from a tutorial written before inference profiles existed.

Diagnosis took about 15 minutes once I stopped guessing. I listed the inference profiles, then confirmed the exact resource the call needed:

```bash
aws bedrock get-inference-profile 
  --inference-profile-identifier us.anthropic.claude-3-5-sonnet-20240620-v1:0 
  --region us-east-1

aws cloudtrail lookup-events 
  --lookup-attributes AttributeKey=EventName,AttributeValue=InvokeModel 
  --max-results 5 --query "Events[].[EventTime,CloudTrailEvent]" --output text


The fix was changing the model ID to the `us.` prefixed inference profile and granting `bedrock:InvokeModel` on both the inference-profile ARN and the underlying foundation-model ARNs in each region the profile routes to. Everything went green on the next deploy. What I took into the exam was the pattern, not the model name: when a managed foundation model won't accept your call, the resource you named and the resource the service resolved are two different things. That's Domain 2 thinking.

A related one that shows up on exam day: `ThrottlingException: Too many requests, please wait before trying again.` On-demand InvokeModel quotas are per model, per region, per account. Getting throttled on one model says nothing about your quota on another, and cross-region inference profiles raise your effective ceiling without a support ticket.

## High-yield traps and 130-minute exam technique

These are the distinctions I saw repeatedly in practice questions:

- Bedrock Guardrails (content filters, denied topics, contextual grounding checks) versus WAF (HTTP-layer rules) versus Macie (S3 data discovery) versus Comprehend (PII detection in text you already have).
- Provisioned throughput buying consistent capacity at a commitment versus on-demand paying per token with a quota ceiling.
- Knowledge Bases managing chunking, sync and retrieval versus you hand-rolling OpenSearch and owning the whole pipeline.
- Compliance questions that hinge on data residency — a cross-region inference profile routes across regions, which is a compliance decision, not just a latency one.
- Ordering items that ask you to sequence a RAG pipeline. Write the steps down on the provided scratch paper before clicking.

Timing: 130 minutes for 65 questions is two minutes each. I did one pass, flagged anything that needed a second read, and finished with 22 minutes left — which is roughly the buffer you want, because matching and ordering items take three or four minutes each. There's no penalty for a wrong answer, so nothing gets submitted blank.

And a budget note for the lab weeks: OpenSearch Serverless bills per OCU-hour and doesn't scale to zero on the older collection types, so a knowledge base you built on day 3 will still be billing on day 30. Set a calendar reminder to tear down lab stacks each Friday. We've cleaned up more than one client's forgotten GenAI sandbox through awsgood.com managed services, and it's

Leave a Reply

Your email address will not be published. Required fields are marked *