ExamSprintLab

AI-300 study resources

Updated September 28, 2026

The best AI-300 resources are free: Microsoft’s skills outline, the official practice assessment, the Azure Machine Learning and Microsoft Foundry documentation, and an Azure subscription you build things in. Add realistic practice questions to find your weak domain. The one thing to be careful with is older material: much of what is labelled AI-300 online was written for DP-100 and covers only the machine learning half.

Start with the official material

The skills outline. Five domains, from 10–15% to 25–30%, split between Azure Machine Learning and Foundry. It names every tool and technique the exam covers: MLflow, registries, endpoints, Bicep, GitHub Actions, provisioned throughput, evaluators, hybrid search and fine-tuning. Read it first and keep it open while you study.

The official practice assessment. It is free and available on Microsoft’s AI Skills Navigator rather than on the usual Learn page; you need to sign in there to launch it. Use it once early to calibrate, and once near the end.

The documentation. For Azure Machine Learning: workspaces and managed networks, data assets, environments, components, registries, jobs and sweeps, MLflow, online and batch endpoints with safe rollout, and model monitoring. For Foundry: resources and projects, deployment types and provisioned throughput, the evaluation SDK and built-in evaluators, tracing, and fine-tuning. For Azure AI Search: hybrid search and the semantic ranker. The how-to guides are current and match what the exam tests.

The exam sandbox. Ten minutes in Microsoft’s demo shows the interface and question types.

Lab work is the differentiator

AI-300 describes things you deploy and run, and asks what is wrong or missing. Six labs cover most of it:

  1. A workspace from code. Deploy a workspace with Bicep from a GitHub Actions workflow that authenticates with OpenID Connect. Then disable public access and see what breaks.
  2. Train and track. A command job with MLflow autologging, a sweep job with early termination, and a pipeline built from two components.
  3. Roll out and roll back. Register the model, deploy it to a managed online endpoint, add a second deployment, mirror traffic to it, split 90/10, and roll back.
  4. Watch for drift. A model monitor with a data drift signal and an alert, then feed it shifted data.
  5. Foundry from code. A Foundry resource and project from Bicep, a model deployment called with a managed identity, and a system prompt kept in Git.
  6. Evaluate, trace, tune. A small RAG app evaluated for groundedness, relevance and safety, traced in Application Insights, then improved by changing chunk size and switching to hybrid search.

Delete the resource group after each session. Compute instances, online endpoints and provisioned deployments cost money while idle.

DP-100 material: use with care

DP-100 courses still teach workspaces, compute, MLflow, sweeps, pipelines and endpoints well, and those topics are in the AI-300 lifecycle and infrastructure domains. Use them for that, and nothing else. They say little about GitHub Actions, Bicep, model monitoring, Foundry deployments, evaluators or fine-tuning, which together are more than half of AI-300. Anything that still says “Azure AI Foundry” was written before the rename to Microsoft Foundry; check it against the current documentation.

Worth paying for

Practice tests written for AI-300, not adapted from DP-100. Their value is diagnosis: they show which domain is weakest while there is still time to fix it. The free 20-question test here is weighted like the real exam and is a good first check.

A small Azure budget. Twenty or thirty dollars of lab time, spent deliberately, teaches more than any video.

What to avoid

Braindumps. They break Microsoft’s exam agreement, risk revocation of every certification you hold, and do not help with scenario questions that describe a situation rather than recite a fact.

Portal click-path memorisation. Menus change, and the Foundry portal has changed often. The exam asks which setting, command, deployment type or evaluator meets a requirement, not where a button is.

Theory-heavy machine learning courses. AI-300 does not test algorithm derivations. Time spent there is time not spent on deployment and evaluation.

A sensible order

  1. Read the skills outline once.
  2. Take the official practice assessment cold, to see where you stand.
  3. Work through the five domains with the documentation and the labs above, following the four-week study plan.
  4. Take the practice test and find your weakest domain.
  5. Go back to that domain’s documentation and repeat its lab.
  6. Book once you score 80% or more on realistic questions twice in a row.

Is this the right exam?

If you operate machine learning or generative AI on Azure, yes. If you build AI applications and agents rather than run them, AI-103 fits better. See the Azure AI certification path for how the exams relate.