AI-103 validates that you can plan, build, secure and operate generative AI and agentic solutions on Microsoft Foundry. It is a decision-heavy developer exam: most questions ask which model, which service, which tool, which deployment type — not "what does this do?". You are expected to read Python and recognise what an SDK snippet does or what is wrong with it.
How to use these notes
Start with Foundry Foundations to build the mental model — the Foundry resource, projects, models, tools and the renamed vocabulary. Then work the five skill pages in outline order, keep the SDK deep dive open while you practise, and drill Exam tips & caveats in the final week. These notes are a companion, not a replacement: pair every section with hands-on time in the Foundry portal and the official Learn modules.
The single biggest thing to get straight before you study
This exam sits on top of a platform-wide rename. Azure AI Foundry is now Microsoft Foundry; Azure AI Services are now Foundry Tools; the Assistants API is now the Responses API; threads, messages, runs and assistants are now conversations, items, responses and agent versions; and the hub + Azure OpenAI + Azure AI Services resource sprawl collapses into a single Foundry resource with child projects. Exam items and documentation use the new names, but plenty of tutorials, blog posts and practice tests still use the old ones. The Foundations page carries the full translation table — learn it first and the rest of the syllabus stops being confusing.
The platform in one picture
Almost every AI-103 question lives somewhere on this diagram. If you can place a service on it, you can usually eliminate two of the four answers.
flowchart TB
APP["Your app or agent code<br/>one project endpoint"] --> PRJ
subgraph PLANE["Wrapped by one governance and observability plane — Entra ID · RBAC · guardrails · private networking · Azure Policy · tracing · evaluations · dashboards"]
subgraph RES["Foundry resource — one Azure resource, one RBAC and network boundary"]
direction TB
PRJ["Foundry project<br/>child resource · your workspace"]
PRJ --> M["Foundry Models<br/>10,000+ models<br/>deployments and routing"]
PRJ --> A["Foundry Agent Service<br/>prompt and hosted agents<br/>Responses API"]
PRJ --> T["Foundry Tools<br/>Content Understanding · Language<br/>Speech · Translator · Vision"]
PRJ --> K["Knowledge and retrieval<br/>Azure AI Search · Foundry IQ<br/>file search · memory"]
end
end Skills measured at a glance
The weighting is not even. Skills 1 and 2 together are roughly 55–65% of the exam — budget your study time accordingly, and do not let the three 10–15% areas eat your schedule.
Suggested study order
flowchart LR F["Foundations<br/>resource, project,<br/>vocabulary"] --> S1["Skill 1<br/>Plan and manage<br/>25–30%"] S1 --> S2["Skill 2<br/>GenAI and agents<br/>30–35%"] S2 --> S3["Skill 3<br/>Computer vision<br/>10–15%"] S2 --> S4["Skill 4<br/>Text and speech<br/>10–15%"] S2 --> S5["Skill 5<br/>Information extraction<br/>10–15%"] S3 --> SDK["SDK deep dive"] S4 --> SDK S5 --> SDK SDK --> TIPS["Exam tips<br/>and caveats"]
Study pages
Foundry Foundations
The mental model everything else hangs off: the Foundry resource and project, Foundry Models, Foundry Tools, connections, the Responses API and the vocabulary shift from Azure AI Foundry.
Start here → Skill 1 · 25–30%Plan and manage an Azure AI solution
Choosing models and services, designing the infrastructure, deployment types, quotas and cost, security and keyless auth, monitoring, and responsible AI guardrails.
Explore → Skill 2 · 30–35%Implement generative AI and agentic solutions
The biggest domain: RAG, prompt agents and hosted agents, tools and toolboxes, multi-agent orchestration, approvals, evaluation, prompt engineering and observability.
Explore → Skill 3 · 10–15%Implement computer vision solutions
Image and video generation, inpainting and mask-based edits, multimodal understanding and captioning, Content Understanding for visuals, and visual responsible AI.
Explore → Skill 4 · 10–15%Implement text analysis solutions
Entity, sentiment and summary extraction with prompts or Azure Language, structured JSON output, translation, and the full speech stack including Voice Live.
Explore → Skill 5 · 10–15%Implement information extraction solutions
Retrieval and grounding pipelines in Azure AI Search, vector/hybrid/semantic ranking, skillsets and integrated vectorization, and document extraction with Content Understanding.
Explore → Deep dive · PythonFoundry SDK deep dive
The Python code the exam makes you read: AIProjectClient, the Responses API, conversations, function tools, streaming, evaluation and tracing — with the classic-to-new migration map.
Drill these → Final reviewExam tips & caveats
Key numbers, decision trees, the naming traps that cost marks, a scenario-to-answer lookup table and a pre-exam checklist for the night before.
Drill these →Before you book
Exam reality check
Expect multiple choice and multiple response, drag-and-drop ordering, "build list" sequencing, hotspot (dropdown) items and yes/no statement sets, plus possible case studies. You will be shown short Python snippets using the Foundry SDK and asked what they do, what to insert, or what is wrong. Most questions cover generally available features; commonly used preview features can appear — and in Foundry a great deal is still in preview, so know which side of the line each capability sits on.
Two habits that pay off more than extra reading
One: for every service, be able to say in one sentence what it does that the neighbouring service does not — Content Understanding vs Document Intelligence, web search vs Grounding with Bing, prompt agent vs hosted agent, file search vs Azure AI Search. The exam is built on those boundaries. Two: build one small RAG agent end to end in a real project. A single hands-on run teaches you the endpoint shapes, role assignments and error messages that four chapters of reading will not.
Disclaimer
These notes are an independent study aid for learning purposes only and are not affiliated with or endorsed by Microsoft. Foundry ships changes almost weekly, and Microsoft renames and re-scopes features frequently — always verify every detail against the official Microsoft documentation linked throughout before relying on it.