Azure AI certification path

Start with AI-901 if you are new to AI on Azure, then move to AI-103 if you build AI apps and agents in code, AI-200 if you build the back-end services they run on, AI-300 if you run models and generative AI in production, or AB-620 if you build agents without writing code. Microsoft 365 administrators who govern Copilot start with AB-900 instead. SC-500 is the security specialisation and assumes you already work in cloud security. There is no need to collect them all.

The order that makes sense

Microsoft’s AI exams are not a single ladder. They fan out from two fundamentals exams into five quite different associate-level roles, and the right route depends on what you actually do.

Where to start

If you have never worked with AI services, take AI-901. It is genuinely a fundamentals exam: it tests whether you can recognise which capability fits a scenario, not whether you can build one. It is one of the shortest exams on this path, and it gives you the vocabulary the associate exams assume you already have.

If you already build things on Azure, skip AI-901. It will teach you little and the credential carries little weight with people who can read a CV. Go straight to AI-103, AI-200, AI-300 or AB-620.

There is no prerequisite enforcement here. Microsoft does not require AI-901 before AI-103. The question is only whether you need it.

AI-103 or AB-620?

These two exams sound similar and are not. The distinction is whether you write code.

AI-103 is for developers. It expects you to work in Python against Azure AI Foundry, understand retrieval-augmented generation properly, and know how to run AI services in production — quotas, identity, monitoring, cost. Roughly two-thirds of the score sits in planning, managing and building generative AI and agentic solutions. It replaced AI-102, which Microsoft retired in June 2026, and it shifted substantially towards agents and RAG in the process.

AB-620 is for people building agents in Copilot Studio. It tests design decisions — scope, knowledge sources, actions, escalation, testing — rather than code. If your work happens through configuration rather than a Python file, this is your exam and AI-103 would be a frustrating detour.

Plenty of people do both, but do them in the order your work demands. Taking the one that matches your job first means the study reinforces what you already do.

AI-103 or AI-200?

Both are developer exams, and they split the application down the middle. AI-103 covers the intelligent part: models, agents and RAG pipelines in Microsoft Foundry. AI-200 covers everything that part runs on and talks to: containers on Container Apps and AKS, vector search in Cosmos DB, PostgreSQL and Redis, Service Bus, Event Grid and Functions, Key Vault and OpenTelemetry. It replaced AZ-204, which Microsoft retired on July 31, 2026, as the associate exam for Azure developers.

If you mostly write prompts and agent logic, take AI-103. If you mostly build APIs, pipelines and data layers, take AI-200. Back-end developers who held AZ-204 should usually start with AI-200.

Where AI-300 fits

AI-300 is the operations exam. It replaced DP-100, which Microsoft retired on 1 June 2026, and it is less about training a good model than about running one: Azure Machine Learning workspaces deployed with Bicep, pipelines in GitHub Actions, model registration and rollout, drift monitoring, and the same discipline applied to generative AI in Foundry — evaluation, tracing and optimisation. Take it if you own the path from notebook to production. Former DP-100 holders will find about half of it familiar.

AB-900 or AB-620?

AB-900 is a fundamentals exam for Microsoft 365 administrators: the core services, data protection and governance for Copilot, and the basic admin tasks for Copilot and agents. AB-620 is for people who build the agents. If you decide who gets Copilot and what it can reach, start with AB-900; if you design agents in Copilot Studio, go to AB-620.

Not AI, but often asked: AZ-900 and AZ-104

AZ-900 covers Azure itself — cloud concepts, core services, governance and cost — and suits anyone who needs the platform vocabulary before AI-901 or an associate exam. AZ-104 is the Azure administrator exam: identity, storage, compute, networking and monitoring. It is not an AI exam, but it is the standard credential for people who run Azure, and the prerequisite for the Azure Solutions Architect Expert certification. Windows Server administrators take AZ-802 instead, which replaced AZ-800 and AZ-801 in September 2026.

Where SC-500 fits

SC-500 is not an AI exam with security on top; it is a cloud security exam that now includes AI workloads. It covers identity and access, securing storage, databases and networking, securing compute, and posture management. The AI content is a slice, not the spine.

Take it if you work in security. If you are a developer hoping to broaden, it is a hard exam to pass on reading alone — it assumes you have configured these controls, argued about them, and cleaned up after them.

How long each takes

Assuming you already work in the area and study consistently:

• AI-901 — two to three weeks
• AI-103 — four to six weeks
• AI-200 — four weeks, more without Azure development experience
• AI-300 — four weeks with Azure Machine Learning experience, six to eight without
• AB-900 — four weeks, six without Microsoft 365 admin experience
• AB-620 — four weeks
• SC-500 — six to eight weeks, longer without security experience

These are honest estimates for people with relevant daily work. Without it, double them and expect to build things rather than read about them.

What to do next

Pick the exam that matches your work, read its exam format guide to see what the real thing actually asks, and take the free 20-question practice test in that section. If you score under fourteen, the domain guides will tell you where the gaps are before you spend anything.

If you are still undecided between the developer and the low-code route, the comparison guide inside each section lays out the difference against the official skills outlines rather than the marketing.

Still choosing between clouds? See AIF-C01 vs AI-901 at foundational level and MLA-C02 vs AI-103 at associate level. Check the exam changes tracker before booking — AI-102 was retired and replaced by AI-103 in June 2026, and AZ-204 by AI-200 in July 2026.

Coming to AI from a business role rather than a technical one? See AI certifications for non-technical roles.