AI-103 study plan: 6-week & 10-week schedules

Here's a concrete, week-by-week plan for Microsoft AI-103 (Developing AI Apps and Agents on Azure), in two tracks: 6 weeks if you're already an Azure developer, an AI-102 holder or already building in Microsoft Foundry, and 10 weeks if you code in Python but are new to Azure. Both run on the same daily routine of about an hour (roughly 6–8 hours a week), both follow the official domain weights, both put hands-on time in Microsoft Foundry into every week, and both finish with timed mocks. Book the exam when you're consistently scoring 80%+ on fresh questions.

AI-103 isn't a "describe" exam; it's a "choose and build" exam. Most questions describe an app or agent and a constraint, and ask which service, deployment type, role, tool or setting fits. The wrong answers are real options that would work for a slightly different need. So this plan pairs Microsoft's free training with practice questions and hands-on work from week one, and ends with four cheat sheets and the Azure CLI setup commands worth knowing.

The exam you're planning for

Generative AI and agents plus planning and managing make up 55–65% of the exam, so both tracks spend most of their weeks there. Because the 700/1000 score is scaled and Microsoft doesn't publish the question count, aim higher than the mock pass line: 80%+ on fresh questions in every domain, not 70%.

Which track are you on?

Those line up with our estimates in how long to study for AI-103, which also has the per-domain hour split.

The daily routine (every study day)

• 20 minutes — learn: the matching module in Microsoft's free self-paced AI-103 training. Read for when you'd choose this over the alternative, not just what the service does.
• 20 minutes — hands-on: do the thing you just read about in Microsoft Foundry: the portal one day, Python with the Foundry SDK the next.
• 15 minutes — practice questions on that day's topic on CrushCert, plus a few mixed questions from earlier weeks so nothing goes cold.
• 5–10 minutes — review misses and flashcards: read the explanation for every wrong answer and say what each wrong choice would be right for; then run flashcards on the lists that blur together (deployment types, roles, agent tools, evaluators).

Track 1: the 6-week plan (Azure developer, AI-102 holder, or Foundry user)

• Week 1 — plan and set up (domain 1): start with a 30-question mixed set to find your gaps. Then choosing models (LLMs, small language models, reasoning, embedding, multimodal) vs Foundry Tools; the Foundry resource and project model; the project endpoint vs the OpenAI v1 endpoint; deployment types (cheat sheet 2); keyless auth with Microsoft Entra ID; and RBAC (cheat sheet 3). Run Set Up a Foundry Resource, Project and Model and Pick the Right Foundry Tool.
• Week 2 — manage, secure and responsible AI (domain 1): quotas in TPM and RPM, HTTP 429 and backoff, Batch vs Provisioned for cost; tracing to Application Insights; private endpoints and agent network isolation; Guardrails and their default behavior, Prompt Shields for user prompt and document attacks, and evaluators (cheat sheet 4). Run Match the Safety Control.
• Week 3 — generative AI apps (domain 2): connect with AIProjectClient and the Responses API; RAG with Azure AI Search and citations; JSON mode vs structured outputs; temperature, top_p and output limits; what reasoning models don't support; reflection loops. Build a small RAG app. Run Build the RAG Pipeline in Order.
• Week 4 — agents (domain 2): prompt vs hosted agents and agent versions; conversations and responses; tools (File Search, Azure AI Search, Code Interpreter, function calling, OpenAPI, MCP, A2A, web search); MCP require_approval and allowed_tools; memory (preview); multi-agent patterns with Microsoft Agent Framework; agent tracing and evaluators. Build one agent with a tool. Run Choose the Agent Tool.
• Week 5 — vision, text and extraction (domains 3–5): gpt-image generation and mask edits, Sora 2 (preview), vision chat for captions and alt text, Content Understanding analyzers; language models vs Azure Language for new text tasks, Azure Translator, Azure Speech and Voice Live; Azure AI Search ingestion, hybrid search and the semantic ranker, skills, Foundry IQ, and Document Intelligence vs Content Understanding (cheat sheet 1). End the week with a 50-question mixed quiz and record your score per domain.
• Week 6 — mocks and the exam: Microsoft's free Practice Assessment, then a full 50-question CrushCert mock on the 120-minute timer. Drill the weakest domain with fresh questions, take a second mock, and book once you're at 80%+.

Track 2: the 10-week plan (Python developer new to Azure)

About an hour a day, six days a week. Use the seventh to catch up or rest.

Weeks 1–3 — plan and manage an Azure AI solution

• Week 1 — set up: get an Azure subscription, open Cloud Shell, create a resource group, then a Foundry resource and project in the Foundry portal (ai.azure.com). Deploy a model with Global Standard and call it from Python using DefaultAzureCredential instead of a key. Run Set Up a Foundry Resource, Project and Model.
• Week 2 — choose and secure: models vs Foundry Tools, endpoints, deployment types (cheat sheet 2), Microsoft Entra ID and RBAC (cheat sheet 3), managed identities, private endpoints. Run Pick the Right Foundry Tool.
• Week 3 — operate and govern: quotas and 429 handling, tracing and monitoring, Guardrails, Prompt Shields, and the evaluators (cheat sheet 4). Run Match the Safety Control.

Weeks 4–7 — generative AI and agentic solutions

• Week 4 — generative apps: the Foundry SDK and Responses API, system messages and few-shot examples, parameters, reasoning models, structured outputs.
• Week 5 — RAG: retrieve with Azure AI Search, add chunks with citations, generate, then evaluate groundedness and relevance. Build it over a few of your own documents. Run Build the RAG Pipeline in Order.
• Week 6 — single agents: create a prompt agent with create_version(), run it in a conversation, add tools, and set MCP approval rules. Run Choose the Agent Tool.
• Week 7 — multi-agent and operations: Microsoft Agent Framework orchestrations (sequential, concurrent, handoff, group chat, Magentic), the A2A tool, memory (preview), agent tracing, reflection loops, and mixing rules with LLM steps.

Weeks 8–10 — vision, text, extraction, mocks

• Week 8 — computer vision (domain 3): gpt-image generation and edits (a mask is a PNG, the same size as the input, with transparent pixels where the edit goes), Sora 2 video (preview), vision chat models, Content Understanding for images and video, and guardrails for multimodal content.
• Week 9 — text analysis and information extraction (domains 4–5): language models for summaries and sentiment in new work, Azure Language for PII and entities, Azure Translator, Azure Speech and Voice Live; Azure AI Search indexers, hybrid queries, the semantic ranker, skills and Foundry IQ; Document Intelligence vs Content Understanding (cheat sheet 1). Finish with a 50-question mixed quiz and record your score per domain.
• Week 10 — review, mocks and the exam: your weakest domain with fresh questions; Microsoft's free Practice Assessment; a full CrushCert mock on the 120-minute timer; drill what it exposed; a second mock. Two in a row at 80%+ and you're ready to book.

Cheat sheet 1: which grounding or retrieval option?

In Azure AI Search, hybrid queries run text (BM25) and vector search in parallel and merge results with Reciprocal Rank Fusion, and the semantic ranker reranks the top 50 results. Its captions and answers are verbatim text from your index; it doesn't generate new text.

Cheat sheet 2: deployment types

Quota is set per subscription, per region, per model and per deployment type, in tokens per minute. Going over it returns HTTP 429; handle it with retries, exponential backoff and jitter.

Cheat sheet 3: Foundry roles

RBAC applies only when you authenticate with Microsoft Entra ID; an API key grants full access without role restrictions. That's why the exam favors keyless access with DefaultAzureCredential and least-privilege roles.

Cheat sheet 4: which evaluator?

Pair evaluators with Guardrails: by default they block medium- and high-severity content in the four harm categories (hate, sexual, violence, self-harm) and turn on Prompt Shields for user prompt attacks. Detection of document (indirect) attacks is off by default, so turn it on when your app reads untrusted documents, emails or web pages.

Azure CLI commands to know

These are the setup commands CrushCert's Azure CLI lab has you type, in order. Learn what each one does and its key flags; the last column is the lab step that drills it.

In code, the model parameter takes the deployment name, not the underlying model name.

The last day before the exam

• No new topics, no full mocks. A 20-question mixed quiz, the flashcard deck, the four cheat sheets and the CLI table above.
• On exam day: you have 120 minutes, and the exam may include interactive components. Read every scenario for the words that decide it ("least privilege," "without keys," "data must stay in the EU," "lowest cost," "new development," "requires approval"), and keep a steady pace so a long scenario doesn't squeeze the last questions.

More on AI-103

Not sure what to practice with? See our honest look at the best AI-103 practice tests, free and paid. Still deciding whether to sit it at all? Read is AI-103 worth it? and how long it takes by background. Everything in the prep is laid out on the AI-103 practice test page.

Keep reading

Frequently asked questions