AI-901 lesson 12: cross-domain scenarios
Most AI-901 questions describe a need in a sentence or two and ask for one of two things: the capability that fits, or the next step in building it. Real scenarios mix several needs, so the skill is splitting them apart. This lesson gives a decision table, three worked scenarios and a routine for reading the text.
Decision table: need, capability, place in Foundry
| The need | Capability | Where it happens |
|---|---|---|
| Answer in a set tone, language or scope every time | System prompt on a deployed model | Playground to test, then the chat client |
| Answer from the company’s own documents | Grounding (retrieval-augmented generation) | Knowledge attached to the prompt or agent |
| Perform steps and call tools toward a goal | Agent with instructions, model and tools | Agent created and tested in the portal, then a client |
| Judge mood, find names or places, pick key phrases, shorten text | Text analysis in Azure Language | Foundry resource, called from a client |
| Convert audio to text or text to audio | Azure Speech recognition or synthesis | Foundry resource, Speech SDK or REST |
| Reason over spoken input in one step | Multimodal model with audio input | Deployed model, tested then called |
| Answer questions about a photo | Multimodal model with image input | Deployed model |
| Create a new picture from a description | Image-generation model | Deployed model, playground |
| Named fields from forms, scans, calls or video | Content Understanding analyzer | Foundry resource, analyzer then client |
Scenario 1: the field-service helper
A 120-person heating company wants technicians to photograph a boiler’s data plate, hear a short spoken summary of the service history, and ask follow-up questions in plain language.
Split it. Reading the plate is an image question; if the plate values must land in fields, that is Content Understanding, and if the technician just asks “what model is this?”, a multimodal model suffices. The spoken summary is speech synthesis, or a model with audio output. The follow-up questions over the company’s manuals need grounding, and the whole assistant has standing rules that belong in the system prompt, such as safety warnings that must always appear.
Scenario 2: the complaints desk
A 500-person insurer records complaint calls. Management wants a transcript, the customer’s mood, the policy numbers mentioned and a two-line recap of each call.
Transcript: speech recognition. Mood: sentiment analysis on the text. Policy numbers: entity detection, or a field in an analyzer. Recap: summarization. Because several of these start from audio, an alternative is a single Content Understanding audio analyzer returning transcript, summary and fields together. Wording decides: separate text techniques lean toward Azure Language, structured results from recordings toward Content Understanding.
Scenario 3: the support assistant
A 70-person software vendor wants a chat assistant that answers only from its help articles, politely declines other topics and, for refund requests, looks up the order. Which parts are prompting, grounding and agent behaviour?
Declining other topics and staying polite are system prompt rules. Answering only from help articles is grounding. Looking up an order is a tool call, which moves the design from a plain chat client to an agent with a tool. The next step the exam might ask about is to create and test that agent in the portal before building a client application for it. Responsible AI also appears: users should know they are talking to AI (transparency), and answers need review for reliability, as covered in the responsible AI principles.
A reading routine
- Find the verb. Classify, summarize, transcribe, describe, create, extract, decide: the verb usually names the capability.
- Find the input and the output. Audio in and text out differs from text in and audio out. Image in and text out differs from text in and image out.
- Find the constraint. Words such as simplest, every conversation, structured fields or without writing code remove options.
- Check the layer. Ask whether the question wants a capability, a prompt setting, a deployment step or a client step.
Key points
- Most questions describe a need and ask for a capability or the next step.
- Multi-part scenarios split into separate capabilities; solve each part before choosing a bundle.
- Standing rules go in the system prompt; company knowledge is grounding; tool use makes an agent.
- Audio, image, text and document inputs each have a typical capability.
- Contrast the pairs that sound alike: recognition vs synthesis, interpretation vs generation, text analysis vs extraction.
- Typical order of work: capability, portal test, lightweight client.
Check yourself
Question 1. A 250-person law firm wants an assistant that always replies formally, answers only from its own policy documents and can file a request in a ticket system. Which design covers all three needs?
- A. An agent with system instructions, grounding knowledge and a ticket tool
- B. An image-generation model with a style prompt
- C. Azure Speech synthesis alone
- D. Sentiment analysis on each question
Show answer
Answer: A
Formal tone is a system prompt rule, answering from policy documents is grounding, and filing a ticket is a tool, which makes the solution an agent. A bare image model, a speech-only design or sentiment analysis addresses none of these needs.
Want more questions like this? Full AI-901 practice tests →
Question 2. A scenario asks for a spoken answer to a spoken question, a written transcript saved for audit, and a consistent brand voice. Which statement is correct?
- A. Keyword extraction handles all three needs
- B. Image generation produces the voice
- C. Azure Speech recognition and synthesis fit the transcript and voice needs
- D. A text-only model alone can accept the audio directly
Show answer
Answer: C
Dedicated speech recognition produces the transcript and speech synthesis supplies a chosen voice, with a model in between for the reply. Keyword extraction does not convert audio, image generation is unrelated, and a text-only model cannot take audio unless something converts it.
Want more questions like this? Full AI-901 practice tests →
Question 3. A scenario states: pull the invoice number and total from thousands of scanned PDFs into a database, with the simplest approach. What is the best reading of it?
- A. Generate images of each invoice
- B. Use a Content Understanding analyzer for documents
- C. Use speech synthesis to read each invoice
- D. Write a custom chat prompt for each PDF
Show answer
Answer: B
The output goes into database columns and the input is scanned documents, which points to a Content Understanding analyzer. Image generation creates pictures, speech synthesis produces audio, and a hand-written prompt per file is heavier and less consistent than a prebuilt analyzer.
Want more questions like this? Full AI-901 practice tests →
This final lesson ties together both exam domains: Identify AI concepts and capabilities (40-45%) and Implement AI solutions by using Microsoft Foundry (55-60%). Reinforce it with the AI-901 cheat sheet and the Foundry implementation recap.