Microsoft AI Certification Study Guide 2026: AI-901, AI-103, AI-300

By Kyle Tice · Updated October 19, 2026 · Facts checked against official exam pages

For most people the Microsoft AI path is now AI-901 (fundamentals), then AI-103 (building AI apps and agents on Microsoft Foundry). Data scientists and ML engineers should take AI-300 (MLOps and GenAIOps) instead of, or after, AI-103. If you work with Microsoft 365 Copilot rather than Azure code, the AB exams are your track.

The old AI-900, AI-102 and DP-100 exams are gone. This guide explains what replaced them, the official skill weights, a study plan for each exam, and how to use Microsoft’s free practice assessments and free renewals.

The June 2026 retirements

Microsoft’s retirement page lists these as retired:

If you already earned one of these, it stays on your Microsoft Learn transcript. You can no longer take the exams or renew the retired associate certifications. Job postings will keep saying “AI-102” for a while; AI-103 is the current equivalent, so say so in your résumé.

Coming up: the AB-730 skills outline changes on October 20, 2026. The new version drops “create and manage Microsoft 365 Copilot agents”, adds a new domain on driving outcomes with agents (prebuilt agents and Microsoft Copilot Cowork), and reduces the weight of the other three domains. If you’re sitting AB-730 after that date, study the new outline. AB-100 and AB-900 also received minor updates on October 14, 2026.

Which Microsoft AI exam to take, and in what order

Prices vary by country. As of October 2026, US prices are about $99 for fundamentals exams and $165 for associate and expert exams. Check the exam page for your region before you budget.

Who should skip: AI-901 adds little if you already code in Python against Azure OpenAI or Foundry. Go straight to AI-103. The AB-730 and AB-731 exams prove fluency, not engineering skill, so don’t expect them to carry weight in a developer interview.

AI-901: Azure AI Fundamentals

AI-901 is not just a renamed AI-900. It now expects you to read basic Python and build lightweight apps in Microsoft Foundry.

Skills measured (as of April 15, 2026)

The first area covers responsible AI principles, how generative models work, model selection and AI workloads (text, speech, vision, information extraction). The second has you deploy a model in Foundry, build a chat client with the Foundry SDK, create a single agent, and use Azure Speech and Azure Content Understanding. Passing score is 700.

3-week plan (about 15–30 hours)

• Week 1: Concepts. Responsible AI’s six principles, generative vs. traditional AI, model parameters (temperature, max tokens), common workloads.
• Week 2: Foundry hands-on. Deploy a model, write a 30-line Python chat client, create and test an agent in the portal.
• Week 3: Speech, vision and Content Understanding labs. Take the free practice assessment until you consistently pass, then book.

AI-103: Azure AI Apps and Agents Developer Associate

Skills measured (as of April 16, 2026)

Expect questions on choosing models and Foundry services, deployment and CI/CD for Foundry projects, quotas and cost, managed identity and private networking, safety filters and evaluators, RAG, multi-agent orchestration, tracing and observability, image and video generation, Content Understanding pipelines and hybrid search. Python is the expected language.

8-week plan (about 80–120 hours)

• Weeks 1–2: Plan and manage. Foundry resource and project setup, model deployment types, quotas, managed identity, keyless auth, content safety configuration.
• Weeks 3–5: Generative AI and agents, the largest area. Build a RAG app, an agent with function calling and memory, then a multi-agent workflow with an approval step. Add tracing and evaluations.
• Week 6: Vision and multimodal: captioning, visual Q&A, image generation and editing, prompt injection via images.
• Week 7: Text, speech and information extraction: structured JSON extraction, translation, speech as an agent modality, Azure AI Search indexing with skills and OCR.
• Week 8: Practice assessment, MeasureUp or other practice tests, and redo weak labs.

AI-300: Machine Learning Operations Engineer Associate

AI-300 replaced DP-100, but it’s a different exam. DP-100 was about training models. AI-300 is about operating them, both classic ML and generative AI.

Skills measured

Tools named in the outline: Azure Machine Learning, Microsoft Foundry, MLflow, GitHub Actions, Bicep and Azure CLI. You’re expected to know entry-level DevOps.

8-week plan (about 80–120 hours)

• Weeks 1–2: Azure ML workspace, datastores, compute, environments, components and registries. Deploy a workspace with Bicep from a GitHub Actions workflow.
• Weeks 3–4: Model lifecycle. MLflow tracking, AutoML, hyperparameter sweeps, pipelines, model registration, real-time and batch endpoints, safe rollout, drift monitoring.
• Week 5: GenAIOps infrastructure. Foundry projects with RBAC and private networking, serverless vs. managed compute deployments, provisioned throughput, prompt versioning in Git.
• Week 6: Evaluation and observability. Groundedness, relevance, coherence and fluency metrics, safety evaluations, tracing, token cost tracking.
• Week 7: RAG tuning (chunk size, thresholds, hybrid search) and fine-tuning with synthetic data.
• Week 8: Practice and review.

The AB exams: Copilot, agents and business roles

AB-100 has a prerequisite. To earn the Expert certification you need AB-100 plus one qualifying associate certification. Microsoft’s list includes AI-103, AI-300, AB-620, PL-400, AB-210, AB-410, AB-250 and several Dynamics 365 associates. Check the current list on Microsoft Learn before you plan.

Study plans for AB exams: AB-730 and AB-731 take one to two weeks for most people who already use Copilot daily. AB-900 takes two to three weeks and leans heavily on Microsoft Purview, Entra ID and SharePoint oversharing controls. AB-100 is an eight-week effort even for experienced architects; the Deploy area (ALM, testing, governance, security for agents) is 40–45% of the exam.

Free official resources first

1. Microsoft Learn study guides. Each exam has one with the current skills outline and a change log. Read it twice: once at the start, once the week before your exam.
2. Free practice assessments. Microsoft now hosts them on AI Skills Navigator (sign-in required). They’re the closest thing to the real question style.
3. Microsoft Learn learning paths and the Exam Readiness Zone videos. Free and maintained by Microsoft.
4. The exam sandbox at aka.ms/examdemo. Try it once so the interface doesn’t surprise you.
5. GitHub repos. microsoft/generative-ai-for-beginners and microsoft/ai-agents-for-beginners are free and hands-on.
6. YouTube. John Savill’s Technical Training covers Azure AI topics clearly and for free. Older AI-900 cram videos are now dated; skip them for AI-901.

Paid resources

• MeasureUp is Microsoft’s official practice test provider. It’s the paid option most worth considering for AI-103 and AI-300.
• Video courses. Udemy has AI-103 and AI-300 courses. Check the “last updated” date and confirm the course covers Microsoft Foundry and agents, not just the old AI-102 content. Search AI-103 courses on Udemy.
• Whizlabs sells Azure practice tests and labs as a second question pool. Browse Whizlabs Azure practice tests.
• Coursera hosts Microsoft-authored professional certificates if you prefer a structured course with graded labs. See Microsoft courses on Coursera.

For AI-901, you don’t need to pay for anything beyond the exam.

Practice-exam strategy

• Use the free practice assessment as a diagnostic on day one, then again at the end. It tells you which skill areas are weak.
• Weight your hours by the skills table. On AI-103, generative AI and agents is up to 35% of the exam. Don’t spend a week on vision if it’s 10–15%.
• Watch for case studies and drag-and-drop ordering. Microsoft exams use more question formats than AWS. The sandbox shows them.
• Expect some preview features. Microsoft says most questions cover GA features but some may cover commonly used preview features. Read the Foundry docs, not just course slides.
• No dumps. They break the Microsoft exam agreement and often contain wrong answers for the newer exams anyway.

Hands-on labs to do

You can do all of these on an Azure free account or a pay-as-you-go subscription with a budget alert.

AI-901

1. Deploy a chat model in Foundry and call it from a short Python script.
2. Build a single agent in the Foundry portal with a knowledge source.
3. Extract fields from an invoice PDF with Content Understanding.

AI-103

1. Build a RAG app over your own documents with Azure AI Search hybrid search, citations and an evaluation run.
2. Build an agent with function calling, memory and a human approval step for one risky tool.
3. Wire a two-agent workflow and add tracing so you can see each step and its token cost.
4. Add content safety filters and test them with a prompt-injection attempt.

AI-300

1. Train a model with MLflow tracking, register it and deploy it to a managed online endpoint with a blue/green rollout.
2. Provision the workspace with Bicep from a GitHub Actions workflow.
3. Set up an automated GenAI evaluation workflow that runs groundedness and safety metrics on every prompt change.

Free renewal: the part people forget

Microsoft associate, expert and specialty certifications (AI-103, AI-300, AB-100) expire every year. You renew by passing a free online assessment on Microsoft Learn, which opens in the six months before expiry. It’s unproctored and you can retake it. Fundamentals certifications like AI-901 don’t expire.

Set a calendar reminder. If you let an associate cert lapse, you have to pay for and pass the full exam again.

Related reading

• AIF-C01 vs AI-901
• AI-103 vs AIP-C01
• Pearson VUE online exam guide
• Vouchers and discounts, including past Microsoft AI Skills Fest offers

FAQ

Is my AI-102 still valid?

It stays on your transcript, but AI-102 retired on June 30, 2026, so you can’t renew it. When it expires, earn AI-103 to hold a current Microsoft AI developer certification.

Is AI-901 harder than AI-900?

Yes. AI-900 was purely conceptual. AI-901 puts 55–60% of the exam on implementing solutions in Microsoft Foundry and expects you to read Python.

Do I need AI-901 before AI-103?

No. There’s no prerequisite. If you already write Python and have used Azure OpenAI or Foundry, start with AI-103.

What happened to DP-100 study material?

Much of it (Azure ML workspaces, MLflow, pipelines) still applies to AI-300’s MLOps half. You’ll need new material for the GenAIOps half: Foundry deployments, evaluations, prompt versioning and RAG tuning.

How much does renewal cost?

Nothing. Renewal assessments on Microsoft Learn are free.