Azure AI Fundamentals (AI-901) practice questions and exam guide

Face recognition features in Foundry Tools sit behind a Limited Access approval mainly because of:

1. Cost overruns from unexpectedly high call volumes
2. Bandwidth limits on sending large images to the service
3. Responsible AI risk around privacy and possible misuse
4. Storage capacity needed to hold the enrolled face images

Answer: C — Responsible AI risk around privacy and possible misuse

C) Right answer - facial recognition raises real privacy and misuse concerns, so Microsoft gates the sensitive features behind an eligibility review. A) Spend is controlled with quotas and budgets, not gating. B) Bandwidth is an engineering matter. D) Storage is cheap and is not why the feature is restricted.

Which statement does NOT describe one of Microsoft's six responsible AI principles?

1. Systems should treat comparable groups of people equitably
2. Systems should perform dependably and safely under expected conditions
3. People should be able to understand how and why a system behaves
4. Systems should minimize the cloud spend incurred by every inference call

Answer: D — Systems should minimize the cloud spend incurred by every inference call

D) Correct — controlling inference cost is an engineering and financial concern, not one of the six responsible AI principles. A) Tempts as a possible omission, but this describes fairness, which is a listed principle. B) Tempts likewise, but this describes reliability and safety. C) Tempts likewise, but this describes transparency.

A large language model (LLM) is best described as:

1. A managed relational database engine for structured rows
2. A foundation model trained on text that generates language
3. A network protocol for passing messages between services
4. A web application framework for building user interfaces

Answer: B — A foundation model trained on text that generates language

B) Right answer - an LLM is a foundation model with a very large parameter count trained on huge text corpora, and it generates human-like language. A) A database stores and queries records rather than generating text. C) A protocol defines message exchange on the wire. D) A framework structures application code and has no learned parameters.

An application sets a low max tokens value and users report that answers stop mid-sentence. What does this parameter actually do?

1. It instructs the model to write more concisely so answers fit within a budget
2. It sets a hard ceiling on the number of tokens generated
3. It limits how many tokens of prompt the deployment will accept per request
4. It reserves a token allowance that is refunded when a response finishes early

Answer: B — It sets a hard ceiling on the number of tokens generated

B) Correct — max tokens is a hard cap on generation, and when it is hit the response simply stops, which is why answers appear truncated. A) Tempts as the single most common misconception; the parameter does not change the model's style, only where generation halts. C) Tempts by reversing the direction; the cap applies to output, while input is bounded by the context window. D) Tempts by inventing a billing mechanic that does not exist.

An operations group wants software that, given a goal such as "resolve this ticket", plans its own steps, calls internal tools to look up and update records, and continues until the goal is met or it escalates. Which workload category does this describe?

1. Text analysis, because the ticket text must be interpreted before acting
2. Information extraction, because record fields are read out of internal systems
3. Agentic AI, because the system plans and takes actions toward a goal autonomously
4. Generative AI, because the resolution notes are written in natural language

Answer: C — Agentic AI, because the system plans and takes actions toward a goal autonomously

C) Correct — agentic AI is the category in which a model-driven system reasons about a goal, decides on a sequence of actions, invokes tools and iterates with limited human intervention. A) Tempts because text is interpreted along the way, but text analysis produces insights from text rather than taking actions. B) Tempts because data is read, but information extraction pulls structured values from documents and does not act on a goal. D) Tempts and is the closest wrong answer: generation is a component, but writing text alone lacks the planning and tool use that define the workload.

Which is NOT a feature of Azure Language in Foundry Tools?

1. Named entity recognition across documents
2. Detection of personally identifiable data
3. Classifying the subject of an image
4. Extractive and abstractive summarization

Answer: C — Classifying the subject of an image

A, B and D) All are text capabilities of Azure Language in Foundry Tools. C) Correct — image classification is handled by Azure Vision in Foundry Tools, not by the language capability.

Azure OpenAI in Foundry Models gives you:

1. Access limited to open-source community models exclusively
2. Managed access to OpenAI models with enterprise controls
3. Unlimited free model usage with no consumption charges
4. Direct low-level control of the underlying GPU hardware

Answer: B — Managed access to OpenAI models with enterprise controls

B) Right answer - the service runs OpenAI models inside Azure with enterprise identity, networking and regional data handling applied. A) These are proprietary OpenAI models; open-weight models sit elsewhere in the model catalog. C) Usage is billed by consumption or reserved capacity. D) The hardware is abstracted away from the customer.

A support tool must transcribe recorded calls and then summarize them. The MOST appropriate approach in Microsoft Foundry is:

1. Deploy an image generation model and prompt it with the audio file
2. Use Azure Speech in Foundry Tools, then a generative model to summarize
3. Use Azure Translator in Foundry Tools to convert the audio into a summary
4. Fine-tune a text model directly on the audio waveform files

Answer: B — Use Azure Speech in Foundry Tools, then a generative model to summarize

B) Correct — speech-to-text produces the transcript, and a generative model then does the language task over that text. A) Image generation cannot process audio. C) Translation converts between languages; it does not transcribe or summarize. D) Fine-tuning on raw waveforms is not how a text model is adapted.

The PRIMARY problem that retrieval-augmented generation addresses is that a model:

1. Cannot produce text in a structured format
2. Does not know private or recent data
3. Is unable to follow a given system prompt
4. Cannot accept more than one prompt example

Answer: B — Does not know private or recent data

B) Correct — the gap RAG closes is knowledge the model was never trained on: internal documents and information newer than training. A) Structured output is a prompting matter. C) Models do follow system prompts. D) Multiple examples are routine.

Embeddings are:

1. Compressed image formats that shrink the storage footprint
2. Numeric vectors for text so similar meanings sit close
3. Point-in-time database backups kept for disaster recovery
4. Compiled machine code emitted by an optimizing build tool

Answer: B — Numeric vectors for text so similar meanings sit close

B) Right answer - an embedding maps text into a vector so that semantically similar passages land near each other, which is what makes vector search and grounding work. A) Image compression is unrelated to meaning. C) Backups are copies of stored data. D) Compiled code is an executable artifact, not a similarity space.

A team submits an image-generation prompt and receives a refusal. Which explanation is MOST likely?

1. Image models must be fine-tuned before producing output
2. A content safety filter blocked the prompt as violating
3. Image generation requires an accompanying audio file
4. A mask is mandatory for every image generation request

Answer: B — A content safety filter blocked the prompt as violating

B) Correct — content safety filters evaluate image prompts and block requests that violate policy, returning a refusal instead of an image. A) A generation model is not required to be fine-tuned first. C) Image models do not need an accompanying audio file. D) Masks are only needed for edits, not for generation.

A product team wants a synthesized narrator whose voice matches their brand's existing spokesperson style. Which capability of Azure Speech in Foundry Tools addresses this?

1. Speaker diarization
2. Language identification
3. Custom neural voice
4. Pronunciation assessment

Answer: C — Custom neural voice

C) Correct — custom neural voice lets an organization build a synthetic voice matched to a chosen speaker, subject to consent and responsible-use review. A) Diarization separates speakers in recognition output. B) Language identification detects the spoken language. D) Pronunciation assessment scores a learner's speech.

What does the max tokens setting limit, and what is the classic mistake?

It caps the length of the generated response, not the length of the prompt. The classic mistake is setting it too low and getting output cut off mid-sentence, then blaming the model — the response is truncated, not finished.

What is a multimodal model?

A model that accepts more than one kind of input — text plus images, and in some cases audio — in a single prompt and reasons over them together. It removes the need to pre-convert an image or a recording into text before the model can use it.

Name the six Microsoft responsible AI principles.

Fairness • Reliability and safety • Privacy and security • Inclusiveness • Transparency • Accountability. The first four are operational qualities of the system; transparency and accountability are governance qualities that sit over the whole lifecycle.

What is Microsoft Foundry?

Microsoft's unified platform for building generative AI solutions — a model catalog, deployment and endpoint management, prompt and agent tooling, evaluation and safety controls, and the Foundry Tools capabilities, all reachable from the Foundry portal or the SDKs.

What is Artificial Intelligence (AI)?

Systems that perform tasks that typically require human intelligence — perception, reasoning, language, learning.

Give two capabilities Azure Vision in Foundry Tools offers that you would not build with a chat prompt.

Optical character recognition tuned for dense printed and handwritten text at scale, and image analysis returning structured tags, captions, object bounding boxes and people detection. These are deterministic, high-volume and priced per image, where a chat prompt is a per-token reasoning call.