Microsoft Certified – Machine Learning Operations Engineer Associate (AI-300)

Learning plan

About 100 hours of focused study, split into 5 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 10 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• A data science background - you can train and evaluate a scikit-learn model in Python
• Basic Git and GitHub Actions, plus enough Azure CLI to create a resource group
• An Azure subscription. Use low-priority or serverless compute and delete clusters after each lab to keep costs down

1. Weeks 1–2 · 18 h

   Azure Machine Learning workspace and infrastructure as code

   • Create a workspace and manage datastores, compute instances, compute clusters and serverless compute
   • Register data assets, custom environments and reusable components, and share them across workspaces with a registry
   • Deploy a workspace and its dependencies with Bicep and the Azure CLI ml extension from a GitHub Actions workflow
   • Configure GitHub federated (OIDC) credentials for Azure, RBAC on the workspace, and a managed virtual network with private endpoints

   Hands-on: Write a Bicep file that creates an Azure Machine Learning workspace with Storage and Key Vault, then deploy it from a GitHub Actions workflow that signs in with OIDC instead of a stored secret.

   Use: AI-300 study guide · AI-300T00 course and self-paced paths · Set up MLOps with GitHub and Azure Machine Learning

2. Weeks 2–5 · 28 h

   Model training, registration and deployment

   • Track experiments with MLflow autologging and compare runs across jobs
   • Run AutoML and a sweep job for hyperparameter tuning, and build a multi-step training pipeline from components
   • Register an MLflow model, run a Responsible AI dashboard evaluation, and archive old model versions
   • Deploy to a managed online endpoint and a batch endpoint, then do a blue-green rollout by shifting traffic and rolling back
   • Set up model monitoring for data drift with an alert that triggers retraining

   Hands-on: Build a pipeline that trains a classifier with MLflow tracking, registers the best model, deploys it to a managed online endpoint as blue, deploys a retrained version as green with 10% traffic, then rolls back to 100% blue.

   Use: MLflow in Azure Machine Learning · Model monitoring in Azure Machine Learning · MicrosoftLearning MLOps labs (GitHub)

3. Weeks 5–7 · 22 h

   GenAIOps infrastructure in Foundry

   • Create Foundry resources and projects with Bicep, managed identities, RBAC and private networking
   • Deploy foundation models as serverless API or managed compute, and know when provisioned throughput units (PTUs) pay off
   • Version model deployments and plan a production upgrade strategy
   • Store prompts in Git, create prompt variants and compare their performance on the same dataset

   Use: Microsoft Foundry documentation · Provisioned throughput concepts · MicrosoftLearning GenAIOps labs (GitHub)

4. Weeks 7–9 · 18 h

   Evaluation, observability and optimization

   • Build a test dataset with column mapping and run groundedness, relevance, coherence and fluency evaluators
   • Run risk and safety evaluations and add an automated evaluation step to a CI workflow
   • Trace a production app and monitor latency, throughput and token cost per request
   • Tune RAG chunk size, overlap, similarity threshold and hybrid search, then A/B test retrieval settings
   • Fine-tune a model with real and synthetic data and take it from evaluation to a production deployment

   Hands-on: Run the Foundry groundedness and relevance evaluators on a small RAG app with a 512-token chunk size, re-index at 1,024 tokens, run them again, and commit both evaluation results and the winning prompt to Git.

   Use: Observability and evaluation in Foundry · Hybrid search overview (Azure AI Search)

5. Weeks 9–10 · 14 h

   Practice assessment and review

   • Take the free AI-300 practice assessment and log misses by skill area, weighting the 25–30% model-lifecycle domain most
   • Redo any lab step you needed docs for, especially CLI v2 YAML for jobs and endpoints
   • Practise looking up CLI and SDK syntax quickly on Microsoft Learn, which is available during role-based exams

   Use: AI Skills Navigator practice assessment · Microsoft exam sandbox

You're ready to book when…

• You can write the CLI v2 YAML for a command job, a pipeline and a managed online deployment without copying it
• You can explain how to detect data drift in a classic model and how to monitor groundedness in a generative app
• You can deploy a workspace or Foundry project end to end from GitHub Actions using OIDC
• You score 80% or higher on the official practice assessment twice

Exam tips

• AI-300 is currently offered in English only, so non-native speakers can request an extra 30 minutes as an accommodation.
• The exam runs 120 minutes and you can open Microsoft Learn during it, but the timer keeps running while you search.
• A 120-minute Microsoft exam is the format that may contain hands-on labs. Microsoft won't say in advance, so read the overview screens when the exam starts.