1.2 Skills Measured, Format, and How This Guide Is Organized

Skills Measured, Format, and How This Guide Is Organized

Quick Answer: As of 5 March 2026, Exam AI-300 measures five domains: design and implement an MLOps infrastructure (15–20%), implement machine learning model lifecycle and operations (25–30%, the largest), design and implement a GenAIOps infrastructure (20–25%), implement generative AI quality assurance and observability (10–15%), and optimize generative AI systems and model performance (10–15%). Most questions cover general availability (GA) features; preview features may appear if they are commonly used. This guide maps every official bullet into 16 chapters (55 sections) covering Azure Machine Learning MLOps, Microsoft Foundry GenAIOps, evaluation, and retrieval-augmented generation (RAG) plus fine-tuning.

Audience Microsoft is testing

The official study guide’s audience profile is the job, not a slogan. As a candidate you should have subject-matter expertise standing up infrastructure for MLOps and GenAIOps solutions on Azure, together called AIOps. You need experience:

• Training, optimizing, deploying, and maintaining traditional machine learning models by using Azure Machine Learning
• Deploying, evaluating, monitoring, and optimizing generative AI applications and agents by using Microsoft Foundry
• A data science background with Python programming
• Entry-level DevOps practices, including GitHub Actions and command-line interfaces (CLIs)
• MLOps knowledge using Machine Learning, Foundry, GitHub Actions, and infrastructure as code (IaC) with Bicep and Azure CLI

Your responsibilities match the five domains listed below. You work with data scientists, DevOps teams, and stakeholders to deliver scalable AI solutions with automation and monitoring.

If you only train models in a notebook, you are under-prepared. If you only click Deploy on a foundation model in a portal and never version prompts or evaluate groundedness, you are under-prepared on the other stack. AI-300 is the exam for the person who has to make both stacks survive production.

The five skills-measured domains (5 March 2026)

The official study guide last updated the skills measured on 2026-03-05. The certification page has received later content refreshes (including an 13 August 2026 update on Microsoft Learn), but the weights in this table are the March 2026 outline. If you sit long after this section’s lastUpdated date, open the live study guide and diff the bullets before you trust a printed copy.

Weights are ranges. They will not add to a neat 100 percent on every form, and they are not a promise that “25–30%” means a published 15 questions. Use them to budget study hours, not to skip a domain.

The bullets under each skill “are intended to illustrate how we are assessing that skill. Related topics may be covered.” That sentence is Microsoft’s warning that a narrowly memorized bullet list is not a shield. If you understand why a managed online endpoint, a batch endpoint, and a Foundry serverless API endpoint exist, you can survive a related wording you have not seen.

Domain 1 — MLOps infrastructure (15–20%)

This is the Azure Machine Learning platform domain: create the workspace and the objects that live in it, then express that platform as code. Expect items that distinguish compute targets from datastores, that assign RBAC rather than sharing keys, that share assets with registries, that lock a workspace behind private networking, and that decide whether a GitHub Actions workflow should deploy a Bicep template or merely trigger a job. If you cannot create a workspace and its associated Storage, Key Vault, and Application Insights resources in your sleep, start Chapters 2–4 before you chase AutoML.

Domain 2 — Model lifecycle and operations (25–30%)

This is the largest slice and the classic MLOps loop: experiment, register, deploy, watch, retrain. MLflow is the tracking and model-format spine. You must distinguish managed online endpoints (synchronous, low latency) from batch endpoints (asynchronous, file-in/file-out), then operate them: test, progressive rollout, safe rollback. Production is not “accuracy on last week’s test set.” Production is data drift, live performance metrics, and retraining or alert triggers when thresholds are exceeded. Responsible AI evaluation and model archiving sit in this domain because a model that cannot be evaluated or retired is not operable.

Domain 3 — GenAIOps infrastructure (20–25%)

Same operational idea, different product. Microsoft Foundry is where you stand up resources and project environments, identities, and network controls, then put foundation models on serverless API endpoints or managed compute, including provisioned throughput units (PTUs) for high-volume workloads. Prompts are production artifacts: design them, create variants, compare performance, and version them in Git. If your Foundry experience is only a playground chat, this domain will feel unfair. It is not unfair; it is operations.

Domain 4 — Quality assurance and observability (10–15%)

Smaller weight, high discrimination. If two candidates both know how to deploy, the one who can wire groundedness versus fluency, turn on risk and safety evaluations for harmful content, and read traces when latency spikes is the one Microsoft is certifying. Token consumption and resource usage are first-class cost metrics, not an afterthought for a finance ticket. Do not skip this domain because the percentage looks small.

Domain 5 — Optimize generative systems (10–15%)

This is not “pick a large model in the catalog.” It is retrieval-augmented generation (RAG) engineering—similarity thresholds, chunk sizes, retrieval strategies, embedding choice, hybrid search (semantic plus keyword)—and advanced fine-tuning with synthetic data through a development-to-production path. Skipping it because it is “only 10–15 percent” is how people fail a scaled exam that still samples every domain.

GA versus preview

Microsoft’s study guide states two notes you should treat as exam law:

• Most questions cover features that are general availability (GA).
• The exam may contain questions on preview features if those features are commonly used.

Do not memorize every private-preview flag. Do not ignore a widely used preview that product docs treat as the default path in 2026. When an item mentions a preview control that is already the mainstream way to evaluate agents or to deploy a model, answer the operations question; do not freeze because the word preview appears in a Learn article.

Exam format you should expect

Microsoft will not identify specific exam formats or question types before you sit, in order to protect exam security. Use the sandbox. The certification page says the exam is proctored and may include interactive components. This guide does not claim that AI-300 currently includes labs. Interactive components and labs are different statements on different pages.

Practice assessments, including the AI Skills Navigator experience linked from the certification page, show style, wording, and difficulty. They are not the real items, and they are not a picture of exam length or complexity (you may see additional question types, case studies, and possibly labs on some Microsoft exams). Taking a practice assessment is not a replacement for training or for experience using the products.

Some Microsoft exams are localized; those updates lag the English version by about eight weeks. AI-300 is offered in English. If an exam is not available in your preferred language, Microsoft’s duration page describes requesting additional time—follow the live accommodations process rather than assuming a blanket extra 30 minutes on this exam.

Microsoft Learn is available on this associate exam. Extra time is not added. The timer continues. It is a lookup tool for problems where you need to confirm something on Learn, not a license to skip studying YAML, evaluator names, or endpoint SKUs.

How this guide is organized

This OpenExamPrep guide is built so every official skills-measured bullet is taught once, in blueprint order, after this introduction.

1. Chapter 1 (this chapter) — exam facts, skills map, candidate profile and tooling.
2. Chapters 2–4 — Domain 1: workspaces, assets, IaC, Git, network isolation.
3. Chapters 5–9 — Domain 2: training, registration, endpoints, monitoring.
4. Chapters 10–12 — Domain 3: Foundry platform, foundation models, prompts.
5. Chapters 13–14 — Domain 4: evaluation, safety, observability.
6. Chapters 15–16 — Domain 5: RAG optimization and fine-tuning.

That is 55 teaching sections across 16 chapters: 3 intro sections plus 52 mapped content sections (after documented clustering and splits of official bullets). Study time should roughly follow weight: Domain 2 first if you are weak on endpoints and MLflow, but do not zero out Domains 4 and 5.

Suggested study sequence

• Confirm the facts in section 1.1 (score, time, retake, renewal) so logistics do not steal attention on test day.
• Stand up an Azure Machine Learning workspace and a Microsoft Foundry project with CLI v2 or Bicep—you need muscle memory for Domain 1 and Domain 3 items.
• Run one MLflow-tracked job, register it, and deploy both an online and a batch endpoint.
• Deploy one foundation model, version a prompt in Git, and run quality plus safety evaluators.
• Tune one RAG pipeline and walk one fine-tuned model from development to production.

Hands-on work beats rereading percentages. The outline tells you what to touch; the remaining chapters tell you how Microsoft expects you to touch it.

Realistic scenario

Priya is a strong Azure Machine Learning pipeline author. She spends four weekends on compute clusters, sweep jobs, and managed online endpoints, then skims Foundry because “GenAI is the small domains.” On the form she meets several items on PTUs, prompt variants, groundedness versus fluency, and hybrid search. Domain 2 strength cannot compensate if Domain 3–5 bars collapse—and Microsoft will not tell her which items she missed. The score report’s bars are a study map for a retake, not a percent-correct decoder.

Common trap

Treating the five percentages as a permission slip to skip a domain, or treating the March 2026 skills outline as if it were last updated on the certification page’s later marketing date. Another format trap: assuming Microsoft Learn access means the exam is open-book enough to skip YAML keys and evaluator names. Learn is a lookup. The timer does not pause. A third trap is inventing an official AI-300 question count or lab list; Microsoft publishes neither.