AI-300 with Microsoft Learn: study guide, practice assessment and labs

The AI-300 study guide on Microsoft Learn (version dated 5 March 2026) contains the candidate profile, the five domains with their weights and the concrete objectives of each, plus a resources table. It is the document that defines the exam: questions are written from those objectives, and Microsoft notes that most cover generally available features, though some may cover preview features if they are commonly used.

How to use it: print the objectives and mark each as done, read or unknown. The unknown ones are your study plan; the read ones, your lab list. We summarize it in the AI-300 outline by domain.

Microsoft offers a free AI-300 practice assessment, hosted on AI Skills Navigator with an upgraded experience; you need to sign in to launch it. It helps you get used to the question style and spot big gaps, but it is a small set and does not replace practice by domain.

The exam sandbox is different: it reproduces the Pearson VUE interface and the question types (single and multiple choice, drag and drop, hotspot, cases) without AI-300 content. It deserves ten minutes a few days before the exam so you do not waste time on exam day learning the controls.

For the first two domains, the Azure Machine Learning documentation covers workspaces, datastores, compute targets, environments, components, registries, MLflow, AutoML, pipelines, managed endpoints and drift monitoring. For the three generative domains, the Microsoft Foundry documentation covers resources and projects, model deployment (serverless and managed compute, provisioned units), evaluations, traces and observability, RAG and fine-tuning.

Tip: read the documentation with the study guide next to it. Each objective in the guide usually maps to one or two documentation pages; going objective by objective keeps you out of sections the exam does not touch.

With an Azure subscription (the free trial is enough to start) you can build a lab that walks the whole exam. First, infrastructure as code: an Azure Machine Learning workspace deployed with Bicep and Azure CLI from a GitHub Actions workflow, with a managed identity and restricted network access. Second, the lifecycle: a small model trained with a component, registered with MLflow, compared against an AutoML run, deployed to a managed endpoint with a second deployment that receives a fraction of the traffic before you roll back.

Third, the generative half: a Foundry resource and project deployed with Bicep, a foundation model on a serverless endpoint, a prompt versioned in Git with two variants, an automated evaluation with quality and safety metrics over a test dataset, and tracing enabled to see latency and token usage. If you also build a minimal RAG and try two chunk sizes, you have touched the fifth domain. Shut down or delete the resources when you finish: compute and endpoints bill by time.

The lab gives you understanding; the questions tell you whether that understanding holds under pressure. After each lab block, take an AI-300 test for that domain and classify every mistake: wrong service, wrong lifecycle stage, confused metric. Save the 120-minute timed mock exams for the end. The full week-by-week plan is in how to pass AI-300.