AI-300 prerequisites: what you need first
AI-300 shows no mandatory prerequisite exam in the Microsoft study guide I checked, yet its audience profile is demanding: a data science background with Python, experience with Azure Machine Learning and Microsoft Foundry, and an entry-level understanding of DevOps. In practice you need to have trained a model, and to be willing to learn GitHub Actions and Bicep. The study plan assumes that starting point.
What Microsoft says
The audience profile expects “subject matter expertise” in setting up infrastructure for MLOps and GenAIOps on Azure, together called AI operations. It asks for experience training, optimising, deploying and maintaining traditional models with Azure Machine Learning, and experience deploying, evaluating, monitoring and optimising generative AI applications and agents with Microsoft Foundry.
It then says you “should have a data science background with experience in Python programming and an entry-level understanding of DevOps practices, including using tools like GitHub Actions and working with command-line interfaces”. Four named toolsets follow: Machine Learning, Foundry, GitHub Actions, and infrastructure as code with Bicep and the Azure CLI. No earlier certification is named.
The background that helps
Python and model training
You should have trained and evaluated at least one model: splitting data, fitting, comparing metrics, saving the result. Familiarity with MLflow-style experiment tracking, hyperparameter tuning and notebooks will make the lifecycle skill area, the largest, feel natural. See model lifecycle explained.
Azure Machine Learning concepts
Know what a workspace, a compute target, a data asset, an environment and an endpoint are. If you have only trained models locally, spend time creating each of these once in a free subscription.
Git and command-line comfort
Branches, pull requests and commits are assumed, as is running Azure CLI commands. Prompts are versioned in Git too, which is a GenAIOps topic.
Basic DevOps and infrastructure as code
You do not need to be a DevOps engineer. Understand what a pipeline does (runs steps automatically on a change) and what infrastructure as code means (describing resources in a file so they can be recreated). Reading a Bicep resource and a workflow step is the target; see MLOps infrastructure explained.
Generative AI basics
Know what a prompt, a token, an embedding and retrieval-augmented generation are. Evaluation concepts such as groundedness and relevance get more attention than model theory.
What you do not need
| Not needed | Why |
|---|---|
| DP-100 | Retired; no earlier exam is required |
| Deep learning research skills | The exam is about operating models, not inventing them |
| Kubernetes expertise | Managed endpoints are the focus |
| A DevOps career | Entry-level understanding is what Microsoft asks for |
| Another certification | None is listed as required |
Self-check
You are ready to start if you can:
- Describe how you would train a model and record its metrics so two runs can be compared.
- Say what a pull request is for.
- Explain in a sentence why a team would describe infrastructure in a file instead of clicking in a portal.
- Explain what data drift means.
- Name one way to tell whether a generative AI answer is grounded in the source documents.
Stuck on two or more? Learn Git and one simple GitHub Actions workflow first; that is the gap that most often costs time.
If you come from a different background
Data scientists usually need extra time for Bicep, workflows and the command line. DevOps engineers usually need extra time for the machine learning lifecycle and for evaluation metrics. Is AI-300 hard? estimates the effort for each route. With no machine learning experience, start with AI-901 and build small projects.
Where it leads
Read the exam format for the mechanics, then AI-300 versus DP-100 if you studied for the old exam. Then use the study plan. For building generative apps see AI-103; for the back end around them, AI-200.