AI-901 Study Guide: Name the Workload by What Goes In and What Comes Out, Then Build It in Foundry
AI-901 (Microsoft Azure AI Fundamentals) has two halves that call for two different habits. The first half asks you to recognise an AI workload or principle from a short story; the trick there is to look at what goes in and what comes out. The second, larger half asks how you would actually build it in Microsoft Foundry, from the model catalog to a few lines of client code. This guide trains both habits in that order.
Study guide · Microsoft Azure AI Fundamentals (AI-901)

The exam
45 minutes, passing score 700 on a scale of 1 to 1,000. Microsoft does not publish a fixed number of questions (this site's mock exam assumes 45), so pace yourself by time. Two skill areas: Identify AI concepts and capabilities (40–45%) and Implement AI solutions by using Microsoft Foundry (55–60%). Even at Fundamentals level, the study guide says you should be familiar with REST APIs, SDKs and CLIs and know basic Python syntax.
Among the 225 AI-901 questions we wrote for this site, Microsoft Foundry itself appears in 30%, followed by Azure Content Understanding in Foundry Tools (15%), Foundry Agent Service and Azure Language (about 12% each), then Azure OpenAI in Foundry Models and Azure Speech (about 11% each). More than one question in five is a troubleshooting question, unusual for a Fundamentals exam: a reply cut off, an image the model cannot fetch, a model name the code cannot find.
Half one: name the workload by input and output
| In → out | Workload and Foundry tool |
|---|---|
| Audio → text | Speech recognition (speech to text), Azure Speech. Text → audio is speech synthesis. |
| Speech in one language → text or speech in another | Speech translation. Plain transcription never changes the language. |
| Existing text → labels, entities, PII, sentiment | Text analysis, Azure Language. |
| Image → one label / boxes / pixel shapes | Image classification / object detection / semantic segmentation. |
| Image + question → text answer | A vision-enabled multimodal chat model. |
| Text → new image | Image generation, the GPT-image models. |
| Invoice, form or call recording → named fields | Information extraction, Azure Content Understanding. OCR alone returns only text. |
| Prompt → new text | Generative AI. If it also uses tools to look things up or act, it is agentic AI. |
Speech recognition, Azure Speech
Speech translation
Text analysis, Azure Language
A vision-enabled multimodal chat model
Information extraction, Azure Content Understanding
Image generation, the GPT-image models
Prompt in, new text out: generative AI (agentic AI if it also uses tools)
Responsible AI from a short story
Microsoft's six principles appear as scenarios. A hiring model that ranks equally skilled people differently by group: fairness. A system that fails in unusual conditions: reliability and safety. Personal data redacted before analysts see it: privacy and security. A product that works for people with disabilities: inclusiveness. Telling users they are talking to an AI, or explaining which factors drove a decision: transparency. A named person who stays answerable: accountability.
How the models work
Text is split into tokens, and a model writes by predicting the next token again and again. Embeddings are vectors that place similar meanings close together, which is how semantic search finds "Recover a forgotten password" for "I can't sign in". The context window must hold the system message, grounding data, history and the reply. Training changes a model's weights; every chat request is inference with fixed weights, so a deployed model does not learn from your conversations. RAG adds private or changing knowledge with citations; fine-tuning changes behaviour, tone or format.
Half two: build it in Foundry, in four steps
- Pick and deploy. Choose a model in the model catalog (its model card shows details such as supported data types and, for select models, benchmarks) and deploy it. The deployment type decides where prompts are processed: Global Standard is Learn's default starting point; Data Zone Standard keeps processing inside a data zone such as the EU.
- Test in the playground. Write the system message (role, scope, format, what to do when unsure), add few-shot examples, change temperature or top_p but not both, and compare up to three deployments side by side. Guardrails (formerly content filters) are assigned to the deployment.
- Call it from code. A client needs three values: the endpoint, a credential (Microsoft Entra ID through
DefaultAzureCredential, or a key kept out of the code) and the deployment name. The Responses API is the newer, stateful choice; Chat Completions needs your code to resend the whole history every time. - Make it an agent if it must act. In Foundry Agent Service an agent is a saved, versioned combination of a model, instructions and tools (file search, code interpreter, web search, function calling, OpenAPI). A conversation holds the history; each response is one run.
Beginner tip: when you deploy your first model, choose custom settings and give the deployment a name different from the model, for example chat-test. Then open the Code tab in the playground and look at what the model parameter contains: your deployment name, not the catalog name. Questions about "model not found" errors test exactly this, and five minutes of doing it yourself fixes it in memory.
Going deeper: why a private image URL fails even inside your own network
When you put an image URL in a prompt, your app does not send the image. The service fetches it from Microsoft's managed infrastructure. So a blob behind a private endpoint, a storage firewall or a VNet-only address cannot be reached, even if your app sits in the same network and can open the URL itself. The fix is to read the file in your code, encode it as base64 and send it as a data URL (data:image/jpeg;base64,...) in the same image part of the message. A question that says "the URL opens fine from our VM but the model cannot read it" is describing this.
Reading what comes back
A Chat Completions reply carries a finish_reason. stop is a normal ending; length means max tokens cut the answer off, so raise the limit or ask for less; content_filter means a guardrail removed content; tool_calls means the model wants your app to run a tool. For long replies, set streaming on so users see text as it arrives.
Text, speech, vision and extraction apps
Azure Language gives fixed, repeatable output, such as PII redaction that code can parse; a prompted model is flexible but varies. Its core capabilities (PII, language detection, named entity recognition, text analytics for health) are recommended for new work, while sentiment, key phrases and summarisation are now listed as legacy. In Azure Speech, single-shot recognition stops at a pause or after at most 15 seconds, so meetings need continuous recognition; SSML controls voice, rate and pronunciation. For images, gpt-image-2 is generally available and dall-e-3 is retired. In Content Understanding, start with a prebuilt analyzer; in a custom schema each field is extracted (copied as written, documents only), generated or classified.
Going deeper: why a long chat eventually breaks
A model has no memory between calls. With Chat Completions your code appends every user message and every reply to the list and sends all of it each time; with the Responses API the service can keep the history for you. Either way, the whole history counts against the context window on every request, together with the system message, any grounding data and the reply itself. A conversation that runs long enough hits the limit, and Learn suggests tracking token counts and dropping the oldest turns. When a stem says "the bot worked for the first twenty questions and then failed", think context window, not quota.
Old names to recognise and avoid
Azure AI Foundry is now Microsoft Foundry; Azure AI services are now Foundry Tools; content filters are now guardrails; the Image Analysis 4.0 API is deprecated and retires on September 25, 2028. A distractor that uses an old name or a retired model is usually wrong for a new solution.
Practise
Practise in the question bank: every question explains each option, right and wrong. Start with the domain you just studied, then take a timed mock exam.
- AI-901 exam page · All 285 AI-901 questions · Timed mock exam · 13-day study plan
- By domain: Identify AI concepts and capabilities · Implement AI solutions by using Microsoft Foundry
Skills measured as listed in Microsoft's AI-901 study guide. Foundry names, model status and retirement dates change often; check the current exam page and Microsoft Learn.