Passing the AI-900 Exam

As I mentioned in my planning for 2026 post, I wanted to get the AI-900 and AI-102 certifications from Microsoft as part of a larger goal of learning more about AI Engineering and being able to deliver Microsoft Training in that area. Good news! Last week I sat the exam for the Azure AI Fundamentals (AI-900) certification and passed. I thought I would share my experience and what I did to prepare.

Preparing for the Exam

Let me start with what knowledge I had prior to studying for the exam. Like everyone else, I have been using AI for a while now and have been tangentially aware of how it all works. I knew that there were different types of models and that they used matrix math to find probabilistic solutions to queries. I had read about neural networks and how training could be supervised or unsupervised. And I was aware that generative AI was used to like generate things, and that the big advancement that led to modern generative models was transformers.

I guess what I’m saying is that I had a lot of general background knowledge regarding AI, just from being a technologist in the 2020s and reading news articles. But that knowledge lacked structure and was fragmented at best. The point of preparing for the exam was to fill in the gaps in my knowledge in a structured format.

What’s Being Tested

You can go read up on the Microsoft Learn site yourself, but here’s a breakdown of the objectives being tested:

Although this is a Microsoft exam, a lot of the content was foundational to machine learning and AI models and not specific to Microsoft products. That being said, you absolutely should be familiar with the various Azure AI solutions and Microsoft Foundry. The study materials and exercises provided by Microsoft will cover all that ground, but I think you should spend a little extra time getting to know the various services, their capabilities, use cases, and how they tie together.

How I Studied

I only used two resources to prepare for the exam. The first was the official, self-led training on the Microsoft Learn site. Over two weeks, I went through all 14 modules in the training and did all the exercises in each module.

Personally, some of the theory introduced early on in the course felt a bit esoteric. The Fundamentals of Machine Learning module gets into the various types of machine learning available, including regression, binary classification, multiclass classification, and clustering. The content gets into the weeds exploring the ways to enhance the predictive capabilities of regression models, like MAE, MSE, and RMSE. While this is interesting to know about, it isn’t really relevant to the certification. You are not going to be asked when it’s appropriate to use the Root Mean Squared Error or the Coeffecient of Determination for a regression plot. That is some deep data science stuff that is outside the scope of the certification. Is it interesting to read about? Yes. Do you need to understand it? No.

I did my best to dedicate about an hour a day to work on modules in the course. My goal was to complete at least one module a day, although that varied depending on the length of the module. I also avoided trying to cram too much into a single day, since I knew I would not retain the information if I did that.

I also made sure to do all the exercises in each module and not rush through them. To get to know the Azure AI services and platform, you really should take your time and explore during each exercise. Don’t limit yourself to the instructions. Being curious helped me fill in some gaps that weren’t covered by the training.

When I finished the course it was time to test my knowledge. The Microsoft Learn site has a sample quiz you can take as many times as you like and the questions are selected at random from a question bank. Each quiz run is composed of 50 questions that cover all the objectives of the course. There’s two things I noticed about the sample quiz:

1. Questions were often repeated in the same quiz run, sometimes one right after another. That seems like a programming issue on Microsoft’s part.
2. Several questions had typos or grammatical issues. It’s pretty clear Microsoft is not prioritizing question quality in the sample quiz.

Given the sorry state of the sample questions, what’s an enterprising student to do? Why, turn to AI of course!

When I was studying for the AZ-104 exam last year, I used VS Code with GitHub Copilot to create a quiz application and prompt for generating test questions. You can find the project here.

The quiz application itself is a simple webpage that loads a question bank from a JSON file you select. Once loaded, you can take the quiz in Practice Mode or Exam Mode. Practice mode simply means that each question is graded as you answer it, and exam mode waits till the end.

To produce the question bank, the repository has a sample prompt that you can tweak as needed. It also includes a script you can run that will generate the necessary prompt for you by asking questions about the exam you’re studying for. Once you have a solid prompt in markdown, you can then use the GenAI model of your choosing to generate the questions. I’ve had good luck with Claude Sonnet 4.5, but any model that supports MCP should work. You really do want MCP available so it can access the official documentation and study materials of the exam you’re preparing for.

You can ask it to generate questions for the entire exam, or focus on specific objectives. The prompt I used for the AI-102 exam is here. I used it twice to generate two questions banks with 50 questions each. What I really appreciate about the quiz is that it provides references for each question, so if I got something wrong I could go read more about it.

The combination of the official Microsoft training and the quiz grader app prepared me to take the exam and pass. I really do recommend using the quiz grader as a supplement. There were several questions it generated that included material or details that were not covered by the course, but did appear on the exam. I can’t tell you exact questions, only that the Microsoft course does have some gaps in its coverage.

Taking the Exam

The exam itself is your standard online, proctored format through PearsonVUE. Despite a lengthy waiting period, the check-in process was pretty painless. I’ve done these exams before so I know how to set up my testing environment to meet their requirements. My preferred location is the kitchen table with the shades drawn and everything electronic moved or facing away from the testing environment. I also took the test during the day when no one else is in the house, thereby avoiding interruption.

I did make the mistake of scheduling the exam around lunchtime, which is probably a pretty busy time for the testing centers. I was 17th in the queue when I joined, and it took about 10-15 minutes for the queue to clear. Bear that in mind if you plan to sit the test and maybe sign-on in a bit early to maximize your time with the exam- you can sign in up to 30 minutes before your exam time.

It took me about 25 minutes to finish the exam and my final score was in the mid-800s. I can’t comment on the question content itself, other than to say that it aligned well with the objectives and the study materials. This is an entry level exam meant for technical and non-technical folks alike, so don’t overthink it.

Next Steps for Me

Taking this exam served two goals:

1. Achieve the AI-900 certification so I can teach the class as an MCT
2. Begin my journey to achieve AI-102 and beyond

The natural next step is to start preparing for the Azure AI Engineer Associate track. I plan to follow the same study process as before, using the materials on the Microsoft Learn site and my quiz grader app for test questions. I may also take the AI-102 prep courses on Pluralsight, since I have free access as an author.

Beyond the AI-102 certification, I am also planning to pursue more AI Engineering knowledge through non-Microsoft material. Marina Wyss pointed me to the AI Engineer for Developers course on datacamp, and that looks pretty interesting. I’m not a developer by trade, but I think I can stumble through the Python enough to check this out.