More study resources

About this exam

AI-200: Developing AI Cloud Solutions on Azure is a role-based credential aimed at developers who build, deploy and operate AI workloads on Azure, rather than at researchers who only train models. It assumes working knowledge of a programming language (usually Python or C#), REST APIs and the Azure portal, and it tests whether you can move an AI capability from a prototype to a governed production service. The exam follows the delivery lifecycle: provisioning and estimating resources, choosing between prebuilt and custom models, wiring retrieval and prompt flows into applications, orchestrating agents with tools and knowledge sources, and handling images, speech, text and documents. Security, identity, cost control, monitoring and responsible AI are not a separate section; they surface inside scenario questions. Preparation works best in layers. Read the concept documentation for each service, then build a small end-to-end project that calls those services from code, then practise justifying one architecture over another out loud. Because the exam leans on judgement, memorising feature lists is weaker preparation than being able to explain trade-offs. This page is independent study material and is not affiliated with or endorsed by Microsoft; domain weightings shown elsewhere are working estimates pending the official skills outline.

What you'll study

Common questions

About AI-200 preparation

What is the exam format?

Expect a mix of multiple choice, multiple response, drag-and-drop ordering or matching items, and case studies in which one scenario carries several linked questions. You answer case-study items using only the information given in the scenario, so read it before the questions. The official certification page lists the current item types and delivery options.

How hard is it?

Moderately hard for working developers and hard for candidates with no hands-on Azure experience. Most items are scenario-based and ask you to choose or sequence an approach, so recognition of service names is not enough. If you have shipped at least one AI-backed application on Azure, the difficulty drops noticeably.

How much does it cost, how long is it, and what score do I need?

Those details change and are set by the certification body, not by study sites. Check the official exam page for the current fee, appointment length, passing score and any regional pricing. Treat numbers quoted in forums or old blog posts as unreliable, since scheduling and scoring details are revised periodically.

What happens if I fail - can I retake it?

Retakes are allowed, but the waiting period and any fee for a repeat attempt are governed by the certification body's policy, which is published on its site. Plan the first attempt when your practice scores are stable rather than treating it as a diagnostic run, because the retake delay can disrupt a deadline.

In what order should I study the domains?

Start with planning and management, since resource provisioning, identity, networking and cost decisions underpin everything else. Then do generative AI and agentic solutions together, because they share prompt, retrieval and orchestration concepts. Finish with vision, natural language and information extraction, which are more self-contained. Allocate time by the official weightings.

Do I need to write code to pass?

Yes. You should be comfortable calling services through an SDK or REST endpoint, handling credentials and configuration, and reading error responses. Portal-only familiarity tends to fail on items about authentication, throttling, batching and deployment configuration. Build something small but complete; a working sample teaches more than reading about one.

What do candidates most often get wrong?

Three recurring problems: memorising service catalogues without learning when to prefer a prebuilt model over a custom one; skipping security, private networking and monitoring because they feel like infrastructure trivia; and rushing multi-part case studies, where one misread constraint invalidates several answers at once. Practise explaining trade-offs, not reciting features.

How do I know when I am ready to book?

You are close when you can sketch an end-to-end solution on a whiteboard - data flow, identity, model choice, evaluation and monitoring - and defend each decision against an alternative. Consistent scores on practice items across all domains, with no domain you consistently avoid, is a better signal than one strong mock run.

Does the certification expire?

Renewal rules are set by the certification body and are published on its site, along with any free online renewal assessment. Check the specific credential page rather than assuming a fixed validity period, because renewal options differ between role-based certifications and can change over time.