AI-300 Domains Explained With Weightings

The short version

AI-300 has five domains. Two of them (infrastructure and the model lifecycle) cover Azure Machine Learning, and three cover generative AI on Microsoft Foundry. The weightings below come from Microsoft’s AI-300 study guide, checked on October 18, 2026. The “traps” and “what to practice” parts are our own advice, not Microsoft statements.

Domain	Weighting
1. Design and implement an MLOps infrastructure	15-20%
2. Implement machine learning model lifecycle and operations	25-30%
3. Design and implement a GenAIOps infrastructure	20-25%
4. Implement generative AI quality assurance and observability	10-15%
5. Optimize generative AI systems and model performance	10-15%

Domain 1: MLOps infrastructure (15-20%)

What it covers. Microsoft lists three groups of skills. First, workspace resources: creating a workspace, datastores, and compute targets, and configuring identity and access. Second, assets: data assets, environments, components, and sharing assets across workspaces with registries. Third, infrastructure as code: GitHub integration with secure access, deploying with Bicep and the Azure CLI, GitHub Actions workflows, restricting network access, and Git for source control.

In practice. You are the person who makes the environment repeatable. A workspace you created by clicking cannot be rebuilt on demand. One you defined in Bicep can.

Typical traps.

Datastore authentication. Identity-based access works for some storage types and not others, and stored secrets are a common reason an answer is wrong.
Mixing up where an asset lives. Registries share assets across workspaces. A workspace-level asset stays in its workspace.
Network questions that sound like one feature but need two, such as a private endpoint plus the right DNS and access setting.

What to practice. Deploy a workspace from Bicep, run it from a GitHub Actions workflow with federated credentials, create a compute cluster from YAML, and register an environment from a base image plus a conda file.

Domain 2: Model lifecycle and operations (25-30%)

What it covers. Training orchestration (MLflow tracking, automated ML, notebooks, hyperparameter tuning, training scripts, distributed training, pipelines, comparing jobs). Registration and versioning (feature retrieval specification, MLflow models, responsible AI evaluation, archiving). Deployment (real-time or batch endpoints, testing and troubleshooting, progressive rollout and rollback). Monitoring (data drift, performance metrics, retraining or alert triggers).

In practice. This is the classic MLOps loop, and it carries the most weight. If you can only deeply learn one domain, make it this one.

Typical traps.

Traffic features that sound alike. A live traffic split changes who gets which response. Mirrored traffic copies requests without changing what clients see.
Online versus batch endpoints and which deployment settings belong to which.
Drift types. Data drift compares input data. Other monitoring signals compare predictions or model quality.
Sweep job details, such as sampling methods and early termination.

What to practice. Run a sweep, build a two-step pipeline, register an MLflow model, create blue and green deployments, shift traffic, roll back, and set up a drift monitor with an alert.

Domain 3: GenAIOps infrastructure (20-25%)

What it covers. Foundry resources and projects; identity with managed identities and role-based access control; network security and private networking; deploying infrastructure with Bicep and the Azure CLI; deploying foundation models by serverless API or managed compute; choosing models for a use case; model versioning and production strategies; provisioned throughput units; and prompt design, variants, and Git version control for prompts.

In practice. This is the Foundry twin of domain 1 plus deployment decisions. Expect constraints such as data residency, cost, and throughput to decide the answer.

Typical traps.

Deployment types. Pay-per-token, regional, and provisioned options differ in cost, capacity, and where data is processed.
Overflow handling for provisioned capacity compared with simply buying more.
Role names. Foundry roles differ in what they let a person build, deploy, or assign.
Model version upgrade behavior.

What to practice. Create a Foundry resource and project from Bicep, assign roles to a test user, deploy a model two ways, and read how provisioned deployments behave under load.

Domain 4: Quality assurance and observability (10-15%)

What it covers. Evaluation: test datasets and data mapping, quality metrics (groundedness, relevance, coherence, fluency), risk and safety evaluations, and automated evaluation workflows with built-in and custom metrics. Observability: continuous monitoring in Foundry, latency, throughput, and response-time metrics, token and cost tracking, and logging, tracing, and debugging.

Typical traps.

Mapping dataset columns to evaluator inputs. The direction of the mapping matters.
Similar-sounding latency metrics. One measures time to the first token, another time to the whole response.
Confusing a quality metric with a safety evaluation.

What to practice. Run an evaluation with a dataset whose field names do not match the evaluator, then open the results and a trace.

Domain 5: Optimizing generative AI systems (10-15%)

What it covers. RAG tuning (similarity thresholds, chunk sizes, retrieval strategies, embedding model selection and fine-tuning, hybrid search that combines semantic and keyword retrieval, relevance metrics, and A/B testing). Fine-tuning (methods, synthetic data, monitoring tuned models, and managing a tuned model through to production).

Typical traps.

Using RAG when the problem is style, or fine-tuning when the problem is missing facts.
Reading loss curves. Falling training loss with rising validation loss points to overfitting.
Changing the embedding model without re-indexing the data.

What to practice. Change chunk size on a small index and compare answers. Read a fine-tuning job’s metrics and decide which setting to change.

More guides for AI-300

Next step

Use the study plan to spread these hours over four or eight weeks. For a first look at the question style, try the sample questions, and when you want more, see the AI-300 practice tests.

Try the AI-300 practice exams