AI-901 implement AI solutions with Foundry explained
Implement AI solutions by using Microsoft Foundry is worth 55–60% of AI-901 — more than the concepts domain and the single biggest reason people who studied for AI-900 fail this exam. It is hands-on. Microsoft’s audience profile assumes Python syntax and familiarity with REST APIs, SDKs and CLIs, and the objectives are written as things you do, not things you describe.
You will not be asked to write a full application. You will be asked which portal step, which SDK call, or which service produces a stated result — and those questions are much easier if you have actually done it once.
Generative apps and agents
The core of the domain:
- Writing effective system and user prompts. The distinction is examinable. A system prompt sets standing behaviour for every turn: tone, language, refusals, scope. A user prompt is one message. Persistent rules belong in the system prompt.
- Deploying a model and using it in the Foundry portal. Know what a deployment is as opposed to a model, and that you interact with it in the portal before any code exists.
- A lightweight chat client with the Foundry SDK. Connect to a project, send a message, read the response.
- Creating and testing a single agent in the portal. One agent, with instructions and at least one tool — not multi-agent orchestration, which belongs to the associate-level exam.
- A lightweight client application for an agent, calling the agent you built rather than the raw model.
The recurring word is lightweight. The exam is checking that you have wired the pieces together, not that you can build production software.
Text and speech
- A small application that performs text analysis.
- Responding to spoken prompts using a deployed multimodal model — the model handles the audio directly.
- Building something with Azure Speech in Foundry Tools.
Keep two routes to speech straight, because questions exploit the overlap: a multimodal model can reason over audio directly, while Azure Speech provides dedicated speech to text and text to speech. Both are valid; which one a question wants depends on whether the emphasis is reasoning or conversion.
Computer vision
- Interpreting visual input in prompts with a deployed multimodal model — you pass an image in and ask about it.
- Creating new visual output with generative models.
- Building a lightweight application with vision capabilities.
The clean split: interpreting an image means a multimodal model consuming it; creating one means an image-generation model producing it. A question describing inspection, checking or answering about a picture wants the first; a question about producing artwork or variations wants the second.
Information extraction
All four objectives here name the same service, Azure Content Understanding in Foundry Tools, applied to different inputs:
- documents and forms
- images
- audio and video
- and a lightweight application built on top of it
If a scenario needs structured fields out of unstructured input — invoice totals, dates, reference numbers — Content Understanding is the expected answer, whatever the input format.
Sample questions
Question 1. Your assistant must always reply in Dutch and must never discuss competitors' products, across every conversation. Where do you configure this?
- A. In every user prompt
- B. In the system prompt
- C. By renaming the model deployment
- D. In the deployment’s capacity settings
Show answer
Answer: B
A system prompt defines standing behaviour that applies to every turn of every conversation, which is what an always/never requirement needs. Repeating it in each user prompt is fragile and easily forgotten, the model’s name carries no behaviour, and deployment capacity settings control throughput rather than instructions.
Question 2. A quality inspector uploads photographs of welds and asks whether each one shows a crack. Which Foundry capability does the application need?
- A. A deployed multimodal model that accepts the image in the prompt
- B. An image-generation model
- C. Azure Speech in Foundry Tools
- D. A small language model with a text-only prompt
Show answer
Answer: A
Interpreting an uploaded image and answering a question about it is a multimodal model consuming visual input in the prompt. An image-generation model creates pictures instead of reading them, Content Understanding targets structured extraction rather than open visual judgement, and speech services handle audio.
Question 3. A team needs the speaker's questions from recorded webinars turned into searchable text, and separately needs structured topics and timestamps extracted from the same recordings. Which pairing is correct?
- A. Image generation, then a system prompt
- B. Content Understanding for the transcript, image generation for the topics
- C. A system prompt for the transcript, Azure Speech for the topics
- D. Azure Speech for the transcript, Content Understanding for the structured extraction
Show answer
Answer: D
Converting the speech to text is a speech capability, while producing structured topics and timestamps from audio and video is what Content Understanding is for. Image generation is unrelated, and a system prompt is an instruction rather than a capability that processes recordings.
What to practise
Open the Foundry portal and spend one session doing all five: deploy a model, chat with it, change the system prompt and watch behaviour change, create a single agent with one tool, and call that agent from a short Python script. Then run one document and one recording through Content Understanding. That is a few hours, and it converts the largest domain on the exam from memorisation into recognition.