You Can Pass This Exam For Free
Choose Your Study Path
You have general programming or data science experience but limited hands-on Azure or MLOps knowledge. You need to build foundational cloud ML skills before tackling operations.
Week 1-2Learn Azure Machine Learning fundamentals: create a workspace, understand compute targets (compute instances, compute clusters, serverless compute), datastores, and data assets. Practice through the Azure portal and CLI
Week 3Study MLflow integration with Azure ML: experiment tracking, model logging, model registry, and how Azure ML workspaces act as MLflow tracking servers. Run a simple training experiment with MLflow tracking
Week 4Learn model training orchestration: training pipelines, components, environments, automated ML (AutoML), hyperparameter tuning, and distributed training concepts
Week 5Study model deployment: managed online endpoints vs batch endpoints, progressive rollout strategies, safe rollback, and endpoint testing and troubleshooting
Week 6Learn production monitoring: data drift detection, prediction drift, model performance metrics, alert triggers, and retraining workflows using Azure Event Hubs or Logic Apps
Week 7Deep dive into GenAIOps: Microsoft Foundry project setup, foundation model deployment (serverless API vs managed compute), provisioned throughput units, prompt versioning with Git
Week 8Study GenAI quality assurance: evaluation metrics (groundedness, relevance, coherence, fluency), risk and safety evaluations, test datasets, and custom evaluation workflows
Week 9Learn RAG optimization: chunk sizes, overlap strategies, hybrid search (semantic + keyword), embedding model selection, and fine-tuning methods including synthetic data generation
Week 10Infrastructure as Code: learn Bicep templates and Azure CLI for deploying ML workspaces, GitHub Actions workflows for automation, and network security configurations. Take full practice exams
Week 11-12Take multiple full mock exams, review all incorrect answers. Focus on weak areas, especially Domain 2 (28% weight) and Domain 3 (24% weight). Schedule your real exam when scoring 75%+
Exam Overview
Format
40-60 questions, 120 minutes. Multiple choice, multiple select, drag-and-drop, and case study scenarios. May include interactive lab components.
Scoring
Scaled score 100-1000. Passing: 700. No penalty for wrong answers — always answer every question even if unsure.
Domains & Weights
- Design and Implement an MLOps Infrastructure18%
- Implement Machine Learning Model Lifecycle and Operations28%
- Design and Implement a GenAIOps Infrastructure24%
- Implement Generative AI Quality Assurance and Observability15%
- Optimize Generative AI Systems and Model Performance15%
Registration
$165 USD. Available at Pearson VUE testing centers or online proctored from home. Exam fee is $165 USD. AI-300 replaces the retiring DP-100 exam.
Topic Priority Table
Not all topics are tested equally. Focus your study time on Tier 1 first, then Tier 2. Tier 3 topics rarely appear — just recognize what they do.
Domain 118% of exam
Design and Implement an MLOps Infrastructure
This domain covers setting up the foundational Azure ML infrastructure for MLOps. You must know how to create and manage workspaces, compute targets, datastores, data assets, environments, and components. Also covers Infrastructure as Code using Bicep and Azure CLI, GitHub integration, and network security for ML workspaces.
Key Topics
Azure ML WorkspaceCompute TargetsDatastoresData AssetsEnvironmentsComponentsRegistriesBicepAzure CLIGitHub Actions
Must-Know Concepts
- How to create and manage Azure ML workspaces including identity and access management (RBAC, managed identities)
- Compute target types: compute instances (development), compute clusters (training), serverless compute, and inference compute — know when to use each
- Datastores connect to Azure storage services (Blob, ADLS, SQL) without exposing credentials. Data assets are versioned references to data in datastores
- Environments encapsulate Python packages and Docker images for reproducible training and deployment. Can be curated or custom, and are versioned
- Components are reusable pipeline building blocks with defined inputs, outputs, code, and environment. Shared across workspaces via registries
- Azure ML Registries enable cross-workspace sharing of models, environments, components, and data assets across an organization
- Deploying workspaces and resources using Bicep templates and Azure CLI commands
- GitHub integration for source control and GitHub Actions workflows for automating resource provisioning
- Network security: restricting workspace access with private endpoints, VNets, and managed network isolation
Common Exam Traps
Datastores are NOT the same as data assets. Datastores define the CONNECTION to storage. Data assets define the specific DATA within a datastore
Compute instances are for interactive development (notebooks). Compute clusters are for training jobs. Do not confuse when each is appropriate
Registries share assets ACROSS workspaces. They are not the same as the workspace-level model registry which is local to a single workspace
Bicep templates are declarative — they define desired state. You do not write imperative provisioning scripts in Bicep. Azure CLI is the imperative alternative
GitHub Actions workflows for ML require secure authentication to Azure. The exam tests GitHub integration configuration, not just workflow YAML
Domain 228% of exam
Implement Machine Learning Model Lifecycle and Operations
The heaviest domain at 28%. Covers the full ML model lifecycle from training through deployment to production monitoring. You must know MLflow tracking, AutoML, hyperparameter tuning, distributed training, model registration, endpoint deployment (online and batch), progressive rollout, data drift detection, and retraining triggers.
Key Topics
MLflowAutoMLTraining PipelinesModel RegistryManaged Online EndpointsBatch EndpointsData Drift MonitoringResponsible AI
Must-Know Concepts
- MLflow experiment tracking: configure tracking URI, log metrics, log artifacts, compare runs across experiments, and use the MLflow UI in Azure ML Studio
- AutoML for exploring optimal models: configure classification, regression, and time-series tasks with automated feature engineering and algorithm selection
- Hyperparameter tuning: configure sweep jobs with search spaces, sampling methods (random, grid, Bayesian), early termination policies, and primary metrics
- Training pipelines: compose multi-step pipelines using components, configure data flow between steps, and schedule pipeline runs
- Distributed training approaches for large models: data parallelism (split data across GPUs) and model parallelism (split model across GPUs)
- MLflow model registration: register models with versioning, package feature retrieval specifications, manage model lifecycle stages (staging, production, archived)
- Responsible AI evaluation: use the Responsible AI Dashboard for fairness, interpretability, error analysis, and causal inference assessment
- Managed online endpoint deployment: deploy models as REST APIs, configure traffic splitting between deployments, implement blue-green and progressive rollout
- Batch endpoint deployment: configure parallel inference on large datasets, manage compute allocation, and troubleshoot batch jobs
- Data drift monitoring: configure monitoring signals (data drift, prediction drift, data quality, feature attribution drift), set thresholds, and configure alert triggers
- Retraining triggers: connect monitoring alerts to Event Hubs, Azure Functions, or Logic Apps to trigger automated retraining pipelines
Common Exam Traps
MLflow models registered in Azure ML can be deployed DIRECTLY to managed endpoints. You do not need to convert them to a different format first
Progressive rollout with managed online endpoints uses traffic SPLITTING — you route a percentage of traffic to the new deployment, not deploy to a subset of instances
Data collection for monitoring is AUTOMATIC with online endpoints but requires MANUAL setup for batch endpoints
AutoML does NOT replace the need for data preparation. It automates model selection and hyperparameter tuning, but data quality still matters
Archiving a model in the registry does NOT delete it. It marks it as no longer recommended for use while preserving the artifact for compliance
Domain 324% of exam
Design and Implement a GenAIOps Infrastructure
This domain covers setting up and managing the infrastructure for generative AI operations using Microsoft Foundry. You must know how to create Foundry environments, configure security (RBAC, managed identities, networking), deploy foundation models (serverless vs managed compute), configure provisioned throughput, and implement prompt versioning with source control.
Key Topics
Microsoft FoundryServerless API EndpointsManaged ComputeProvisioned Throughput UnitsBicep TemplatesRBACManaged IdentitiesPrompt Versioning
Must-Know Concepts
- Creating and configuring Microsoft Foundry resources and project environments, including hub-and-project architecture
- Identity and access management: managed identities for secure authentication, RBAC for granular permissions on Foundry resources
- Network security: private endpoints, private networking configurations, and VNet integration for Foundry environments
- Infrastructure as Code: deploying Foundry resources using Bicep templates and Azure CLI
- Foundation model deployment options: serverless API endpoints (pay-as-you-go, no GPU management) vs managed compute (dedicated resources, more control)
- Selecting appropriate models for specific use cases from the Foundry model catalog
- Model versioning and production deployment strategies for foundation models
- Provisioned throughput units (PTUs) for guaranteeing consistent performance on high-volume workloads
- Prompt design and development: creating prompts, building prompt variants, and comparing performance across variants
- Prompt version control using Git repositories for tracking changes and enabling collaboration
Common Exam Traps
Serverless API endpoints support global, data zone, and regional deployment scopes. Global Standard routes across worldwide infrastructure; Data Zone restricts to a geographic boundary; Regional pins to a specific region for compliance
PTUs guarantee throughput but require upfront commitment. Pay-as-you-go serverless is more flexible but does not guarantee capacity during high demand
Microsoft Foundry is the current platform name — it replaced Azure AI Studio and Azure AI Foundry (classic). The exam references Microsoft Foundry; documentation may still use the older name in some places
Prompt versioning with Git is about tracking PROMPT changes, not model changes. Model versioning and prompt versioning are separate concerns
Managed compute deployments integrate with the full MLOps lifecycle. Serverless deployments are simpler but offer less pipeline integration
Domain 415% of exam
Implement Generative AI Quality Assurance and Observability
This domain covers evaluating and monitoring generative AI applications in production. You must know how to create test datasets, implement AI quality metrics (groundedness, relevance, coherence, fluency), configure risk and safety evaluations, set up automated evaluation workflows, and implement comprehensive observability including performance, cost, and debugging capabilities.
Key Topics
Evaluation MetricsTest DatasetsRisk and Safety EvaluationsObservabilityTracingToken MonitoringLatency Tracking
Must-Know Concepts
- Creating test datasets and data mapping for comprehensive evaluation of GenAI applications and agents
- AI quality metrics: groundedness (factual accuracy based on source data), relevance (response addresses the query), coherence (logical consistency and flow), fluency (natural language quality)
- Risk and safety evaluations: detecting harmful content, bias, and policy violations in model outputs
- Automated evaluation workflows using built-in metrics and custom evaluation metrics
- Continuous monitoring in Microsoft Foundry: setting up dashboards, configuring alerts, and tracking trends
- Performance metrics: latency (time to first token, total response time), throughput (requests per second), and response time distribution
- Cost metrics: token consumption (input tokens, output tokens), resource usage, and cost allocation across projects
- Logging, tracing, and debugging: capturing detailed request/response logs, distributed tracing for multi-step GenAI applications, and debugging production issues
Common Exam Traps
Groundedness measures whether the response is factually supported by the provided SOURCE DATA, not general knowledge. A fluent and coherent response can still be ungrounded
Evaluation metrics can be computed on BOTH test datasets and production traffic. The exam tests whether you know when to use each approach
Token consumption tracks BOTH input tokens (prompt) and output tokens (response). Cost optimization may require reducing either or both
Tracing is especially important for multi-step GenAI applications (agents, RAG pipelines) where you need to identify which step caused an issue
Safety evaluations are separate from quality evaluations. A high-quality response can still be unsafe, and vice versa
Domain 515% of exam
Optimize Generative AI Systems and Model Performance
This domain covers optimizing both RAG systems and foundation models for production performance. You must know how to tune retrieval parameters (chunk sizes, similarity thresholds, search strategies), select and fine-tune embedding models, implement hybrid search, evaluate RAG quality, and implement advanced fine-tuning with synthetic data generation.
Key Topics
RAG OptimizationChunk Size TuningHybrid SearchEmbedding ModelsFine-TuningSynthetic DataA/B Testing
Must-Know Concepts
- RAG retrieval optimization: tuning similarity thresholds to balance precision and recall, adjusting chunk sizes and overlap for optimal context windows
- Chunk size strategies: smaller chunks (e.g., 512 tokens) provide more precise retrieval, larger chunks provide more context but dilute relevance. Overlap prevents information loss at boundaries
- Embedding model selection: choosing models optimized for your domain, fine-tuning embeddings for domain-specific vocabulary and concepts
- Hybrid search: combining semantic search (vector-based) with keyword search (BM25) to capture both conceptual similarity and exact matches
- Evaluating RAG performance: using relevance metrics, A/B testing frameworks, and systematic parameter sweeps to find optimal configurations
- Fine-tuning methods: supervised fine-tuning on labeled data, parameter-efficient fine-tuning (LoRA, QLoRA), and instruction tuning
- Synthetic data generation: creating training data for fine-tuning when real labeled data is scarce, using LLMs to generate diverse examples
- Monitoring fine-tuned model performance: comparing against base models, tracking quality degradation, and managing fine-tuned models through the development-to-production lifecycle
Common Exam Traps
Larger chunk sizes do NOT always improve RAG quality. They increase context but dilute retrieval precision. The optimal size depends on query patterns and data characteristics
Hybrid search is almost always better than pure vector or pure keyword search alone. The exam expects you to know this
Fine-tuning changes model BEHAVIOR (style, format, task specialization). RAG adds KNOWLEDGE. If you need to add new facts, use RAG. If you need to change how the model responds, use fine-tuning
Synthetic data for fine-tuning must be diverse and representative. Low-quality synthetic data can degrade model performance rather than improve it
A/B testing for RAG configurations requires holding the LLM constant while varying retrieval parameters to isolate the impact of retrieval changes
Services and Concepts You Must Not Confuse
These pairs appear on nearly every exam. Learn the difference and you'll avoid the most common traps.
Top Mistakes to Avoid
Confusing datastores (connection to storage) with data assets (versioned references to specific data) — they are separate Azure ML concepts
Mixing up managed online endpoints (real-time REST APIs) with batch endpoints (offline bulk inference) — they have different deployment patterns and monitoring capabilities
Thinking serverless API endpoints and managed compute deployments are interchangeable — serverless is pay-as-you-go with no GPU management, managed compute gives more control but requires infrastructure planning
Confusing data drift (input distribution changes) with prediction drift (output distribution changes) — both are monitoring signals but detect different problems
Assuming RAG and fine-tuning serve the same purpose — RAG adds knowledge at query time without changing the model, fine-tuning modifies the model's behavior permanently
Not understanding the difference between Azure ML Environments (software dependencies) and Components (reusable pipeline steps) — environments define WHAT is installed, components define WHAT code runs
Thinking Bicep templates and GitHub Actions serve the same role — Bicep defines infrastructure declaratively, GitHub Actions automates WHEN and HOW to deploy it
Confusing groundedness (supported by source data) with relevance (addresses the query) in GenAI evaluation — a response can be relevant but ungrounded, or grounded but irrelevant
Forgetting that provisioned throughput units (PTUs) are needed for guaranteed performance on serverless endpoints — pay-as-you-go does not guarantee capacity
Assuming AutoML eliminates the need for data preparation and feature engineering — AutoML automates model selection and tuning, not data quality
Exam-Ready Checklist
Can explain all 5 exam domains and their relative weights (18%, 28%, 24%, 15%, 15%)
Know how to create and configure Azure ML workspaces including compute targets, datastores, data assets, environments, and components
Can deploy ML infrastructure using Bicep templates and automate with GitHub Actions workflows
Understand MLflow integration with Azure ML: experiment tracking, model logging, model registration, and deployment
Can configure AutoML, hyperparameter tuning, and training pipelines for model development
Know the difference between managed online endpoints and batch endpoints, including progressive rollout strategies
Can configure data drift monitoring, set alert thresholds, and connect to retraining triggers via Event Hubs or Logic Apps
Understand Microsoft Foundry architecture: project setup, model deployment options (serverless vs managed compute), and PTUs
Can implement prompt versioning with Git and compare prompt variant performance
Know all four GenAI quality metrics (groundedness, relevance, coherence, fluency) and when each applies
Can configure observability including latency tracking, token consumption monitoring, distributed tracing, and cost metrics
Understand RAG optimization: chunk sizes, overlap strategies, similarity thresholds, hybrid search, and embedding model selection
Can explain when to use fine-tuning vs RAG and how to create synthetic data for fine-tuning
Scored 70%+ on at least two full mock exams (700/1000 passing score)
Recommended Resources
Free & Official Resources
Paid Courses & Practice Exams
These are recommended if you prefer a structured learning path. They can save time but are not required to pass.