Microsoft Foundry Essentials

The single biggest AI-900 → AI-901 change: Microsoft Foundry (Azure AI Foundry) is now the center of the exam — the implementation domain (55–60% of questions) is framed around building AI applications and agents on it. You've used the concepts everywhere; this lesson gives you Microsoft's names for them.

What Foundry is

Microsoft Foundry is the unified platform for building, deploying, and managing AI applications and agents on Azure — one portal/SDK surface that brings together model access, your data, orchestration, evaluation, and safety controls. Where AI-900-era Azure treated each AI capability as a separate service you stitched together, Foundry is the opinionated "build it here" layer on top.

The building blocks (exam vocabulary)

• Hubs and projects — the organizational containers: a hub carries shared configuration, security, and resources; projects live under it per application/team. (Governance instincts apply: the hub is where enterprise control lives.)
• Model catalog — the menu of available models: Azure OpenAI models (GPT-series), plus open and partner models (Meta, Mistral, etc.), comparable and deployable from one place. Know that model choice is a catalog action, not a separate procurement per model.
• Deployments — a model made callable: you deploy a model from the catalog to get an endpoint your app calls. Managed capacity/quotas live here.
• Playgrounds — in-portal chat/completions surfaces for prototyping prompts against a deployment before writing code.
• Grounding / "on your data" / RAG — Foundry's first-class pattern for connecting a deployment to your own content (typically via Azure AI Search) so answers cite your documents. This is the productized version of the RAG pipeline you've drilled — the exam will describe it in Foundry terms.
• Agent service — building agents: a model plus instructions plus tools/actions (functions, APIs, search) the agent can invoke, with the runtime handling the tool-call loop you know mechanically from Module 12 lesson 8.
• Prompt flow — visual/code orchestration of multi-step LLM pipelines (prompt → retrieval → model → post-processing), with testing and deployment of the flow itself.
• Evaluations — built-in evaluation tooling: run models/flows against test data and score quality (groundedness, relevance, safety) before shipping. Your "no eval, no deploy" doctrine, as a product feature.
• Content safety integration — Azure AI Content Safety filters wired into deployments to screen inputs/outputs for harmful content.

Copilot fundamentals (also new on the exam)

Two meanings to keep straight:

• Microsoft Copilot products — ready-made assistants embedded in Microsoft 365, GitHub, Windows, etc. You use them; you don't build them.
• Custom copilots / agents you build — your own assistant over your own data and tools, built on Foundry (grounding + agent service + deployments). The exam wants you to know which is which and that Foundry (plus Copilot Studio for low-code cases) is the build path.

Python familiarity

AI-901 expects reading familiarity with Python-flavored SDK snippets — recognize what a few lines do (create a client, call a deployment, pass a prompt), not write them. Your Python exposure from the Lyra platform more than covers it.

Mapping to what you've built

Everything here is a managed version of something you've done by hand: Foundry deployments ≈ your provider-agnostic AI endpoint config; grounding-on-your-data ≈ ainews's RAG; the agent service ≈ the tool-call loop; evaluations ≈ your gold-set doctrine; content safety ≈ your PromptGuard layer. In the exam — and in the interview — you can honestly say the platform formalizes patterns you already operate.