The AWS Generative AI Professional Exam: What the Blueprint Doesn’t Reveal
Tips about the AWS Generative AI Professional exam
Series: Golden Jacket Field Notes (22 articles)
- …
- 17The AWS Generative AI Professional Exam: What the Blueprint Doesn’t Reveal This article
Disclaimer
Originally published on Medium as Week 8 of Golden Jacket Field Notes**
Canonical URL: https://medium.com/aws-tip/the-aws-generative-ai-professional-exam-what-the-blueprint-doesnt-reveal-66b09500af96
This is a republication on AWS Builder Center for the AWS community. The piece below appeared first on Medium and is preserved here in its original form.
— -
Golden Jacket Field Notes. Week 8.
— -
Ten seconds
That is what I had left when I clicked Submit on the AWS Certified Generative AI Developer — Professional exam. No time to revisit a single flagged question. No time to second guess a single answer. About four hours of continuous focus including a thirty minute English as a Second Language (ESL) extension, and my cursor was hovering on the final question with the timer closing in on zero.
I have sat for every active AWS certification exam. I hold Golden Jacket status. I came out of the Advanced Networking Specialty thinking it was the deepest technical exam AWS produced. I was wrong. The Generative AI Professional exam went further into its domain than Networking Specialty went into networking. That is not a sentence I write lightly.
This is my field debrief. What the public blueprint reveals, what it does not, and what you actually face on exam day if you go for AIP-C01.
I sat for the early adopter window and earned the early adopter badge. Everything below respects the AWS Certification Non-Disclosure Agreement (NDA). I will not describe specific questions, scenarios, or wording. I will describe the shape of the exam, the depth it demands, and the path I walked to pass it.
The shape of the exam
AWS Certified Generative AI Developer — Professional, exam code AIP-C01. I sat for the early adopter beta window. The beta carried a longer question set and a longer time allotment than what current general availability candidates face, so verify the official exam guide for the current numbers before you book. The current general availability exam includes 65 scored questions plus 10 unscored experimental items that AWS uses to evaluate questions for future forms, for a total of 75 you sit through.
What has not changed across beta and general availability is the structure. Multiple choice and multiple response items, scenario based, with the unscored questions mixed in and not identified. Scoring runs 100 to 1000 with a passing score of 750. Pass or fail, no partial credit by domain. The exam uses a compensatory model, meaning your total score determines the result.
The job role AWS targets is a Generative AI developer with two or more years of general cloud experience and at least one year of hands-on Generative AI work. The recommended prior certifications are the AI Practitioner foundational, the Solutions Architect Associate, the Machine Learning Engineer Associate, and the Data Engineer Associate. AWS does not gate the exam behind those, but the blueprint assumes you carry the knowledge they validate.
The five content domains and their weights are public:
- Domain 1, Foundation Model Integration, Data Management, and Compliance, at 31 percent
- Domain 2, Implementation and Integration, at 26 percent
- Domain 3, AI Safety, Security, and Governance, at 20 percent
- Domain 4, Operational Efficiency and Optimization for Generative AI Applications, at 12 percent
- Domain 5, Testing, Validation, and Troubleshooting, at 11 percent
If you stop reading the blueprint at the weights, you walk in unprepared. The weights tell you where to spend study time. They do not tell you what AWS expects you to do inside each domain.
The depth claim, defended
When I say this exam goes deeper than Networking Specialty, I am not making a feeling claim. I am describing what AWS asks you to do at the skill level of the blueprint.
Open the official exam guide and read the skill statements. Skill 1.5.5 expects you to design sophisticated query handling systems for Foundation Model (FM) augmentation, including query expansion, query decomposition, and orchestration across services. Skill 3.3.4 expects you to implement continuous monitoring and governance controls including bias drift monitoring, token-level redaction, response logging, and Artificial Intelligence (AI) output policy filters. These are not multiple choice questions about service definitions. These are architecture and implementation problems.
Let me put a personal stake behind that comparison. I came to AWS from a networking background. I am Cisco Certified Network Associate (CCNA) and I spent years in network engineering before I migrated to solutions architecture. Networking is a domain I knew before I ever logged into the AWS console. Even with that background, the Networking Specialty exam was hard for me. I respected what it asked me to know. I prepared seriously. I did not breeze through it. The Generative AI Professional exam was harder than that. That is the comparison I am making, and I am making it from inside both domains.
Networking Specialty asked you to know Border Gateway Protocol (BGP) routing behavior, Direct Connect failure modes, transit gateway design, and hybrid connectivity edge cases. Hard. Specific. Bounded by a mature domain.
The Generative AI Professional exam asks you to integrate Foundation Models, design Retrieval Augmented Generation (RAG) pipelines, configure vector stores, orchestrate agents, govern outputs, evaluate models, manage cost, observe in production, and troubleshoot when something behaves badly. The domain is younger, the services are newer, and the patterns are still being formed. The exam compensates by demanding that you understand not just what the services do, but how they compose under real constraints.
That is the depth claim, and it survives the comparison.
What the blueprint does tell you
The blueprint is honest about scope. Bedrock is the center of gravity. The in-scope service list reads like a tour of the Bedrock ecosystem: Foundation Models, Knowledge Bases, Agents, Guardrails, Prompt Management, Prompt Flows, Data Automation, Titan embeddings, reranker models, and the Bedrock Application Programming Interface (API) surface itself. Around that center, the blueprint lists OpenSearch Service with vector search, Aurora with the pgvector extension, Amazon Q Business and Amazon Q Developer, AWS Lambda, AWS Step Functions, Amazon API Gateway, AWS AppConfig, AWS Glue Data Catalog, AWS CloudTrail, AWS X-Ray, and the AWS Well-Architected Framework Generative AI Lens.
It also tells you what is out of scope. Foundation Model training from scratch is out. Custom hardware for Machine Learning (ML) training is out. Pure data science work disconnected from application integration is out. The exam does not test you on whether you can build a transformer. It tests you on whether you can put one to work in a system that survives production.
The blueprint also discloses question composition. Multiple choice and multiple response, scenario based. No drag and drop. No simulations. Just text, options, and the ten unscored items mixed in that you will not be able to identify.
What the blueprint does not tell you
This is where the field notes start.
First, the questions assume you have run these services. Not read about them. Run them. The difference is the difference between knowing that Bedrock Knowledge Bases support fixed, hierarchical, semantic, and custom Lambda chunking strategies, and knowing what each one actually costs you in retrieval quality when you run it against a real corpus. The exam tests the second kind of knowledge.
Second, the integration depth is the trap. Most questions are not single-service questions. They are composition questions. You are given a constraint set and asked which combination of services produces the desired behavior. Bedrock plus OpenSearch plus Lambda plus Step Functions plus an evaluation harness, all configured a particular way. If your mental model of these services is shallow, you cannot reason about how they fit together under pressure.
Third, the safety and governance domain is heavier than its 20 percent weight suggests, because safety language threads through the other four domains. You can be answering a Domain 2 implementation question and the correct answer hinges on a guardrail behavior or a token-level redaction strategy that lives conceptually in Domain 3. The domains are not silos. They are perspectives on the same system.
Fourth, the exam is unforgiving on partial knowledge. In other AWS exams you can often eliminate two answers on service knowledge and pick the better of the remaining two on architectural intuition. In this exam, two of the four answers are usually defensible, and the difference between them comes down to a specific behavior of a specific service in a specific scenario. If you have not used the service, you cannot tell.
Fifth, evaluation is treated as a first class engineering practice. The exam expects you to know how to evaluate Foundation Model outputs systematically, how to detect drift, how to design ground truth, how to use Bedrock evaluation tooling and how to design custom evaluation pipelines. Most candidates underestimate this. I almost did.
Sixth, observability for Generative AI is not generic observability. The exam expects you to know what to instrument and where the signals come from. Token usage, prompt and response logging, latency distributions across model providers, retrieval relevance scores, and policy violation rates. CloudWatch and X-Ray show up, but the question is what you put through them, not whether you can configure them.
Less SageMaker than you would expect, more Bedrock than you can imagine
This was the surprise for me, and it is the surprise I hear from every candidate I have spoken with.
I came out of the Machine Learning Engineer Associate exam very heavy on Amazon SageMaker. SageMaker Pipelines, SageMaker Studio, training jobs, hyperparameter tuning, deployment endpoints, model registry, monitoring. That exam is a SageMaker exam in practice.
The Generative AI Professional exam is not. SageMaker JumpStart appears in the blueprint and shows up where it makes sense, mostly around alternative Foundation Model hosting and fine tuning paths. But the gravitational center of this exam is Bedrock, and inside Bedrock the depth required spans the full feature surface. Knowledge Bases internals. Agent design and Action Group orchestration. Guardrail policies and their interaction with model outputs. Prompt Management and Prompt Flows. Bedrock Data Automation for unstructured ingestion. Titan and partner embedding model behavior. Reranker model placement in retrieval pipelines.
If you are coming from Machine Learning Engineer Associate expecting Generative AI Professional to be its bigger sibling, recalibrate. They overlap less than the names suggest.
The study path that worked
I am going to be direct about the materials I used because the question I get most often is not what the exam covers. It is how to prepare.
Frank Kane’s video course was my first pass. I used it for breadth and to make sure I had no blind spots in the official scope. Frank’s strength is that he stays close to AWS source material and does not invent depth where the documentation does not support it. For an early adopter exam where the community is still figuring out what to emphasize, that discipline matters.
Stéphane Maarek’s content was my second pass. Stéphane’s strength is exam frame and pacing. He explains how AWS writes questions, where the misdirection lives, and how to read a four hundred word scenario quickly without losing the constraint that actually drives the answer. If Frank gave me the surface area, Stéphane gave me the test taking instinct.
Vladimir’s practice exams on Udemy were the calibrator. I do not use practice exams to learn content. I use them to discover what I do not know. Every miss became a research target. If I missed a question on Bedrock Guardrails behavior, I went to the Bedrock documentation, read the relevant section, then opened the console and ran the configuration myself.
That last sentence is the entire game. The thing that separated this preparation from every other AWS exam I have taken is what came after the courses and the practice exams.
Labs are not optional
Four to five hours per day, almost three months, in the AWS console and in my own AWS account. Not watching. Building.
I built Knowledge Bases against dummy documents and tuned the chunking strategies until I understood what each one cost me. I built agents with Action Groups and watched them invoke the Lambda functions. I configured Guardrails and probed them with adversarial prompts to see exactly where they fired and where they did not. I deployed embeddings models and ran retrieval against OpenSearch and against Aurora pgvector to feel the latency and cost difference. I instrumented prompt and response logging through CloudWatch and looked at the actual data shape. I broke things on purpose, then fixed them.
This is the part that took the most time and the part that was non negotiable. The exam does not reward people who watched videos. It rewards people who ran the services. I have written before in this series about Bedrock Knowledge Bases in production and about the agentic stack with Strands, AgentCore, Model Context Protocol (MCP), and Agent2Agent (A2A). Those articles came out of this same hands on time. The lab work paid double. It got me through the exam and it gave me material to write about.
If you are studying for AIP-C01 and you are not running the services, no course will save you. I am being plain about this because I think the certification industry sometimes oversells courses and undersells lab time. For this exam in particular, the ratio is wrong. You need lab hours, and you need a lot of them.
Time management and the ESL extension
The full standard window plus the thirty minute ESL extension comes to almost four hours of continuous focus. Most candidates do not realize that focus is the constraint, not knowledge.
My pacing strategy was simple in design and brutal in execution. Read the question once for shape. Read it a second time for the constraint. Eliminate two answers fast. Decide between the remaining two by going back to the constraint. If I could not decide in roughly two and a half minutes, flag and move. Do not lose the run trying to win one battle.
That strategy assumed I would have review time at the end. I did not. I came back to the flagged questions in the final pass and the timer ate me. Ten seconds. I was lucky those answers had been correct on the first read.
If you qualify for the ESL accommodation and English is not your first language, take it. The thirty minutes is real, the cognitive cost of processing scenario language at high speed is real, and the questions are long enough that the extension genuinely changes the experience. There is no penalty and no marker on your record. If you would benefit, request it.
A practical note. Drink water before, not during. Eat something with protein an hour before you sit. The exam is long enough that energy management is part of preparation. This sounds obvious. People still skip it.
What this exam teaches you that you take anywhere
I want to close on something that goes beyond AWS.
The blueprint is AWS specific. The services on it are AWS services. The acronyms are AWS acronyms. But the way the exam forces you to think is portable. If you study seriously for AIP-C01, you come out the other side knowing how to:
Reason about Foundation Model selection across capability, latency, and cost. Design RAG architectures that survive bad input. Build evaluation systems that catch drift before users do. Apply guardrails and governance without strangling the application. Instrument observability that surfaces the right signals for Generative AI workloads. Compose agents that do useful work in production, not in demos.
None of that is AWS only. You will recognize the same problems on Azure with Azure AI Foundry, on Google Cloud with Vertex AI, on platforms that run open source Foundation Models on their own infrastructure. The vocabulary changes. The architectural problems do not. AWS happens to be where I work and where I built this knowledge. The knowledge itself moves with you.
That is the strongest argument I can make for this exam, and it is the argument I would make even if AWS did not pay any of my bills. The certification validates AWS specific skill, which is what the badge says it does. But the preparation forces you into a Generative AI literacy that is genuinely transferable. That is rare in cloud certifications, and it is the reason I think this exam matters beyond the AWS ecosystem.
Caveats and disclosures
A few things worth saying explicitly.
I sat for the early adopter window. Question forms and emphasis can shift over time. The blueprint can be revised. AWS occasionally rebalances domains in subsequent versions of an exam. Treat my observations as a snapshot, not a permanent map. Verify the current blueprint when you prepare.
I respected the NDA in this article. If anything reads like it pushes the line, I would rewrite it. Nothing here describes specific questions, scenarios, or wording from the exam.
Your mileage will vary based on your Generative AI background, your hands on time, and your test taking instincts. The path I walked is one path. It is not the only one. If you have stronger ML fundamentals than I had going in, you will spend less time on the foundational layer and more on the integration layer. If you have weaker hands on Bedrock experience, you will need more lab time than I did, not less.
I am not affiliated with Frank Kane, Stéphane Maarek, or Vladimir Raykov. I paid for their content. I name them because they helped me, and the question about study materials is the most common one I get. If something else works better for you, use it.
The takeaway
The AWS Certified Generative AI Developer — Professional exam is the deepest exam in the AWS portfolio for the territory it covers. The blueprint tells you the perimeter. It does not tell you the depth. The depth comes from running the services, breaking them, and understanding how they compose under real production constraints.
If you are serious about Generative AI on AWS, sit for it. Prepare like the technical exam it is, not the certificate it produces. Lab hours. Composition thinking. Evaluation discipline. Time management. And if English is not your first language, take the ESL extension.
Ten seconds. That is what I had left, and that is how close it can be.
Now I am going back to work, building.
This is Week 8 of Golden Jacket Field Notes, a weekly series where I document what I learn building and architecting on AWS. Previous weeks covered Bedrock Knowledge Bases in production, the AWS agentic stack with Strands, AgentCore, Model Context Protocol, and Agent2Agent, payment authorization architecture for credit cards and Pix, MCP context management with Kiro, the Frontier Agents dissection of AWS Security Agent and AWS DevOps Agent, and deterministic-first GenAI governance. I hold all active AWS certifications, including the AWS Certified Generative AI Developer — Professional early adopter badge.
Follow me on Medium for the weekly series: https://medium.com/@erickmancz Connect on LinkedIn: https://linkedin.com/in/erick-mancz
Series: Golden Jacket Field Notes (22 articles)
- …
- 17The AWS Generative AI Professional Exam: What the Blueprint Doesn’t Reveal This article
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article