ExamSprintLab

AI-300 study plan: pass in 4 weeks

Updated September 28, 2026

Four weeks is realistic for AI-300 if you already train models in Python, have used Azure Machine Learning, and can study 8–10 hours a week. Build the infrastructure first, then run the machine learning lifecycle on it, then do the same for generative AI in Foundry, and finish with evaluation and optimisation. Without Azure Machine Learning experience, or without any GitHub Actions background, plan on six to eight weeks.

Before you start

  • Read the official skills outline end to end. It lists every objective in five short sections.
  • Create an Azure subscription with a budget alert. Compute clusters scale to zero, but compute instances, managed online endpoints and provisioned deployments bill while they exist. Delete resource groups at the end of each session.
  • Install the Azure CLI with the ml extension, Python with the azure-ai-ml and mlflow packages, and create a GitHub repository you can push workflows to.

Week 1: MLOps infrastructure (15–20%)

Goal: a workspace you can recreate from code, reached securely from GitHub.

  • Workspace resources: create a workspace and look at what comes with it (storage, Key Vault, Application Insights, container registry). Register a datastore; compare compute instances, compute clusters and serverless compute.
  • Identity and access: assign built-in roles such as AzureML Data Scientist, and give compute a managed identity to read data.
  • Assets: data assets (uri_file, uri_folder, mltable), curated and custom environments, and components defined in YAML. Create a registry and share an environment through it.
  • Infrastructure as code: deploy the workspace with a Bicep file and az deployment group create.
  • GitHub Actions: authenticate with OpenID Connect and a federated credential instead of a stored secret, then run the Bicep deployment from a workflow.
  • Network: understand managed virtual network isolation modes and private endpoints, and what disabling public network access breaks.

Lab: delete your workspace and recreate it by pushing to your repository.

Week 2: model lifecycle and operations (25–30%)

Goal: take one model from experiment to monitored production. This is the largest domain, so give it the most hours.

  • Training: a command job running a script, MLflow autologging, a sweep job with a sampling method and early termination policy, one AutoML run, and a pipeline built from components. Compare runs in the studio.
  • Distributed training: know when to use a distribution setting on a job and what the instance count controls.
  • Registration: register an MLflow model, version it, add tags, and archive an old version.
  • Responsible AI: generate a Responsible AI dashboard for the model and know what each component shows.
  • Deployment: a managed online endpoint with two deployments, a traffic split between them, and a batch endpoint for scoring files.
  • Monitoring: set up model monitoring with a data drift signal and an alert.

Lab: a GitHub Actions workflow that trains, registers and deploys a new version with 10% traffic, then promotes it or rolls it back.

Week 3: GenAIOps infrastructure (20–25%)

Goal: the same discipline for generative AI in Foundry.

  • Foundry resources and projects: create them with Bicep, connect them to other services, and grant access with managed identities and RBAC.
  • Networking: private endpoints and disabling public access on the Foundry resource.
  • Model deployment: compare serverless API deployment with managed compute; understand standard and provisioned throughput deployments and when PTUs pay off; pin model versions and plan upgrades.
  • Model choice: match models to tasks on capability, context length, latency and cost.
  • Prompts: write prompts as files, create variants, compare them on the same test set, and keep them in Git.

Lab: deploy a chat model from code, store its system prompt in your repository, and change it only through a pull request.

Week 4: quality, observability, optimisation (20–30%) and review

Goal: prove the generative AI app is good, see what it costs, and make it better.

  • Evaluation: build a test dataset, map its columns to evaluator inputs, run groundedness, relevance, coherence and fluency evaluators, then risk and safety evaluators. Add an evaluation step to your workflow.
  • Observability: enable tracing to Application Insights, and watch latency, throughput, token use and cost in Foundry monitoring.
  • RAG tuning: vary chunk size and overlap, top-k and similarity thresholds; compare vector, keyword and hybrid search in Azure AI Search.
  • Fine-tuning: know what supervised and preference-based fine-tuning are for, how synthetic data is created and checked, and how a fine-tuned model is deployed and monitored.

Review: take the free 20-question practice test mid-week. Spend the remaining days on whichever domain scored lowest, using its guide in this section.

If you held DP-100

Compress week 2 to a few evenings; most of it is familiar. Put the saved time into week 1, where GitHub Actions and Bicep are new for most data scientists, and into week 3.

How to know you are ready

  • You can explain how a model moves from a dev workspace to production without being retrained.
  • You can say which evaluator measures groundedness and which measures relevance, and what data each needs.
  • You score 80% or more on realistic practice questions twice in a row, on different question sets.

Then book it. See the exam format for the mechanics and the exam day guide for the day itself.