AI-300 Study Guide: Machine Learning Operations Engineer Associate
New certification: AI-300 replaces DP-100 and broadens the scope well beyond data science. It covers both classical MLOps (Azure Machine Learning) and GenAIOps (Microsoft Foundry, generative AI evaluation, fine-tuning). Beta exam opened April 2026, goes live May 2026.
Overview
The AI-300 exam is aimed at engineers responsible for taking AI and ML systems to production and keeping them running reliably. This includes infrastructure-as-code for ML workspaces, CI/CD pipelines for model retraining, deploying and managing large language models on Foundry, and implementing observability for both classical models and generative AI applications.
Exam Details
| Detail | Information |
|---|---|
| Exam code | AI-300 |
| Certification | Machine Learning Operations Engineer Associate |
| Level | Associate |
| Passing score | 700 / 1000 |
| Duration | 120 minutes |
| Cost | $165 USD (varies by region) |
| Beta availability | April 2026 |
| Exam goes live | May 2026 |
| Predecessor | DP-100 (retires June 1, 2026) |
Skills Measured
Design and Implement MLOps Infrastructure (15–20%)
- Create and manage Azure Machine Learning workspaces using Infrastructure as Code (Bicep, Terraform, ARM)
- Configure compute clusters, compute instances, and serverless compute
- Set up data stores and data assets with versioning
- Implement RBAC, managed identity, and network isolation for ML workspaces
- Design CI/CD pipelines for ML workflows using GitHub Actions or Azure DevOps
- Implement environment management and dependency tracking
Implement Machine Learning Model Lifecycle and Operations (25–30%)
- Orchestrate model training using Azure ML pipelines and components
- Implement model registration and versioning with MLflow
- Deploy models to managed online endpoints (real-time inference)
- Deploy models to batch endpoints (asynchronous scoring)
- Implement blue/green and canary deployment strategies
- Monitor deployed models for data drift and performance degradation
- Configure automated model retraining triggers
Design and Implement GenAIOps Infrastructure (20–25%)
- Implement Foundry hub and project configuration using IaC
- Deploy and manage foundation models (GPT-4o, Phi, Llama) for production workloads
- Configure provisioned throughput units (PTU) vs serverless deployments
- Implement prompt versioning and management with source control
- Manage model endpoints, quota, and rate limits
Implement Generative AI Quality Assurance and Observability (10–15%)
- Configure evaluation runs in Microsoft Foundry using built-in evaluators
- Implement custom evaluators for domain-specific quality metrics
- Assess groundedness, relevance, coherence, and safety of AI outputs
- Implement distributed tracing and logging for AI applications
- Set up dashboards and alerts for generative AI observability
Optimize Generative AI Systems and Model Performance (10–15%)
- Optimize RAG pipeline accuracy (chunking, embedding models, reranking)
- Implement fine-tuning workflows for foundation models
- Apply distillation and model customization techniques
- Optimize inference cost and latency (batching, caching, model size selection)
Exam Domain Weights
| Domain | Weight | Study Part | Focus Area |
|---|---|---|---|
| 1 — Design and Implement MLOps Infrastructure | 15–20% | Part 1 | Workspace IaC, RBAC, network isolation, compute, data assets, MLflow, CI/CD |
| 2 — Implement ML Model Lifecycle and Operations | 25–30% | Part 2 | Pipeline components, registry, online/batch endpoints, monitoring, deployment strategies |
| 3 — Design and Implement GenAIOps Infrastructure | 20–25% | Part 3 | Foundry Hub/Project IaC, PTU vs serverless, prompt versioning, APIM gateway |
| 4 — Implement GenAI Quality Assurance and Observability | 10–15% | Part 4 | Built-in evaluators, custom evaluators, CI/CD quality gates, OpenTelemetry tracing |
| 5 — Optimize Generative AI Systems and Model Performance | 10–15% | Part 5 | RAG optimisation, fine-tuning, Batch API, semantic caching, model routing, streaming |
Study priority: Domain 2 (ML Lifecycle) carries the most weight at 25–30%. Allocate study time proportionally — Domains 1 and 2 together account for 40–50% of the exam. Domain 3 (GenAIOps) has grown significantly to 20–25%, so don't underweight it. Domains 4 and 5 together are 20–30% and share many conceptual overlaps around evaluation and optimisation.
MLOps vs GenAIOps — Side by Side
DP-100 vs AI-300 — Key Differences
| Area | DP-100 | AI-300 |
|---|---|---|
| Primary role | Data Scientist (build and train models) | MLOps Engineer (operate models in production) |
| IaC / CI/CD | Not covered | Core domain — Bicep, GitHub Actions |
| Generative AI | Not covered | Full GenAIOps domain on Foundry |
| Monitoring | Basic data drift detection | Full observability — tracing, dashboards, alerts |
| Model training | Heavy focus (AutoML, sweep jobs) | Pipeline orchestration, not hands-on training |
| Evaluation | MLflow metrics on classical models | Both classical metrics and gen AI evaluators |
| Fine-tuning | Not covered | Foundation model fine-tuning and distillation |
Key Tools and Technologies
- Azure ML Python SDK v2 — primary SDK for workspace operations, pipelines, and endpoints
- MLflow — experiment tracking, model registry, model signatures
- Azure AI Foundry SDK — hub/project management, model deployments, evaluations
- GitHub Actions / Azure DevOps — CI/CD pipelines for model retraining and deployment
- Bicep / Terraform — IaC for workspace and Foundry infrastructure
- OpenTelemetry — distributed tracing standard used by Azure for AI observability
- Azure Monitor / Application Insights — monitoring dashboards and alert rules
Recommended Study Path
- Azure ML Pipelines overview — pipelines, components, scheduling
- Deploy models to online endpoints — blue/green, traffic splitting
- MLOps and GenAIOps for AI workloads on Azure — Well-Architected Framework guidance
- Evaluate generative AI apps in Foundry — evaluators, metrics
- Official AI-300 study guide — Microsoft Learn