AI-901 Microsoft Azure AI Fundamentals (Updated Version) Questions and Answers

Providing users with clear information about an AI system’s purpose, functionality, and limitations is an example of the Microsoft responsible AI principle of transparency . Transparency means users should be able to understand when and how AI is being used, what the system can do, and what its limitations are.

In Microsoft Foundry Agent Service, tool_choice controls whether and how an agent uses tools during a run.

Statement 1: Yes

Setting tool_choice to auto lets the agent/model decide whether to call a tool. This statement is true.

Statement 2: No

Setting tool_choice to none does not mean the model decides whether to call a tool. It means the model will not call a tool. This statement is false.

Statement 3: Yes

Setting tool_choice to required means the agent must call one or more tools during the run. This statement is true.

Microsoft documentation describes the tool_choice values as follows: auto allows the model to decide whether to call tools, none prevents tool calls, and required forces one or more tool calls.

The correct answer is endpoint of the model . After a vision-enabled GPT model is deployed in Microsoft Foundry, the deployment exposes an inference endpoint that applications can call programmatically. The application sends requests containing text and, when required, image content to the deployed model through the appropriate chat-completions API endpoint. Microsoft’s documentation explicitly states that to use vision-enabled models, you call the Chat Completion API against a supported model that has been deployed.

The Foundry playground is primarily an interactive environment for testing prompts, model behavior, and multimodal inputs during development. It is not the production interface that an external application uses to submit inference requests. An evaluation pipeline is used to assess model or application quality against defined metrics and datasets, while a training dataset is associated with model training or fine-tuning rather than serving inference traffic.

Microsoft Foundry provides the endpoint and required authentication information after deployment. Applications then invoke that endpoint using an SDK or REST API, passing text and image inputs as part of the request payload.

Statement 1: Voice Live returns only transcribed text. = No

Voice Live is not limited to returning transcribed text. It is designed for real-time spoken conversational experiences and can include speech input, model reasoning, and spoken output.

Statement 2: Voice Live requires you to separately implement speech to text and text to speech services. = No

Voice Live combines the conversational voice pipeline, so you do not need to separately build independent speech-to-text and text-to-speech services for the same real-time voice experience.

Statement 3: Voice Live combines speech to text, reasoning, and text to speech into a single conversational experience. = Yes

This statement is correct. Voice Live is used for real-time conversational AI experiences where spoken input is processed, the model reasons over the request, and spoken output can be returned.

Verifying that machine learning models do not show racial or gender bias is an example of the Microsoft responsible AI principle of fairness .

Fairness means AI systems should treat people equitably and avoid unfair bias across demographic groups such as race, gender, age, or other protected characteristics.

Why the other options are incorrect:

A. Safety focuses on avoiding harm and operating safely.

B. Privacy and security focuses on protecting data and access.

C. Reliability focuses on consistent and dependable operation.

D. Fairness directly addresses bias and equitable treatment.

The Model catalog is used for comparing and deploying a wide range of models for generative AI development.

Microsoft states that the model catalog in Foundry portal is the hub for discovering and using a wide range of models to build generative AI applications. Microsoft training also states that the model catalog is used to compare models by benchmark metrics and deploy models to endpoints.

The requirement is to convert written customer notifications into natural-sounding spoken audio . This is speech synthesis , also known as text to speech .

Microsoft’s Azure Speech documentation describes text to speech as a capability that converts text into natural-sounding synthesized speech. Therefore, for playing written notifications over a phone system, the correct Azure Speech capability is speech synthesis .

Why the other options are incorrect:

A. speaker recognition identifies or verifies speakers by voice.

C. speech recognition converts spoken audio into text.

D. speech translation translates spoken audio between languages.

Statement 1: Generating a response to a user prompt occurs during the inference stage. = Yes

When a deployed generative AI model receives a user prompt and generates an output, that process is called inference . The model is being used to produce a response, not being trained.

Statement 2: A generative AI model generates responses by copying stored documents directly from the model’s training data. = No

Generative AI models do not normally respond by simply copying stored documents from training data. They generate outputs based on learned statistical patterns. In some cases, models may reproduce memorized content, but that is not the normal or intended generation mechanism.

Statement 3: A generative AI model produces output by predicting the next token based on patterns learned from the model’s training data. = Yes

This is correct. Large language models generate text by predicting likely next tokens based on the prompt and learned patterns from training.

AI systems must consistently operate as intended, even under unexpected conditions. = Reliability and safety

This description matches the Microsoft responsible AI principle of reliability and safety , which focuses on ensuring that AI systems behave consistently, safely, and as expected, including under unusual or unexpected conditions.

AI systems must protect and secure personal and business information. = Privacy and security

This description matches the Microsoft responsible AI principle of privacy and security , which focuses on protecting personal, confidential, and business-sensitive information and preventing unauthorized access or misuse.

To initialize the Azure Speech SDK for Python, you create a SpeechConfig object with information for the Speech resource. Microsoft documentation states that SpeechConfig includes information such as the Speech resource key and associated region, endpoint, host, or authorization token . It also shows the Python pattern:

speech_config = speechsdk.SpeechConfig(subscription= " YourSpeechKey " , region= " YourSpeechRegion " )

or, for endpoint-based configuration:

speech_config = speechsdk.SpeechConfig(subscription=os.environ.get( " SPEECH_KEY " ), endpoint=os.environ.get( " ENDPOINT " ))

Therefore, among the provided options, the correct answer is D. the Foundry resource key and endpoint .

Options A, B, and C are incorrect because Speech SDK initialization does not use an OpenAI model deployment name, storage account credentials, or the Foundry project name/project endpoint.

The completed code is:

poller = client.begin_analyze(

analyzer_id= " invoice " ,

input_url=url

)

result = poller.result()

Azure Content Understanding analysis uses a long-running operation pattern. The Python SDK returns a poller from begin_analyze(), and Microsoft documentation states that the SDK poller handles polling automatically when you call .result() .

Therefore, to retrieve the analysis results after processing completes, the correct option is:

result

The other options are incorrect because status checks operation state, wait waits without returning the final analysis object, and get_results is not the method shown for retrieving the begin_analyze() result in this code pattern.

The Model catalog is used for comparing and deploying a wide range of models for generative AI development in Microsoft Foundry.

Microsoft documentation states that the model catalog in Foundry portal is the hub for discovering and using a wide range of models to build generative AI applications . It also includes many models across providers such as Azure OpenAI, Mistral, Meta, Cohere, NVIDIA, and Hugging Face.

The Microsoft Learn module for Foundry models also states that the model catalog is used to explore and filter models , compare models using benchmark metrics , and deploy a model to an endpoint .

For thousands of recorded support calls stored as audio files in Azure Storage, the correct capability is speech to text batch transcription .

Microsoft states that batch transcription is designed to transcribe a large amount of audio data in storage , including audio files in Azure Blob Storage, and that files can be processed concurrently to reduce turnaround time.

Real-time transcription is for live audio, not large stored batches. Text to speech converts text into audio. Speech translation translates speech between languages, but the requirement is to generate transcripts.

The correct answer is natural language processing . In this scenario, a user communicates with a chatbot using ordinary human language, such as asking, “Will it rain today?” The chatbot must interpret the meaning and intent of the user ' s text, process the request, and formulate a natural-language response containing the relevant weather information. These are characteristic natural language processing (NLP) functions.

The current AI-901 study guide explicitly includes text analysis among the common AI workloads candidates must identify, and it also covers techniques for processing and interpreting human language. Microsoft documentation on conversational language understanding explains that conversational systems interpret user utterances to determine the user ' s intent and extract relevant information before the application performs an appropriate action.

Although the weather information itself may originate from a forecasting system, the scenario being tested centers on the chatbot interaction —understanding a natural-language question and responding conversationally. Therefore, prediction and forecasting is not the best classification of the complete interaction. Computer vision is also incorrect because no images or visual information are being analyzed.

AudioOutputConfig(filename= " output.wav " )

The question specifically states the application must save the synthesized audio to a file . In the Azure Speech SDK for Python, speechsdk.audio.AudioOutputConfig(filename= " output.wav " ) directs the synthesizer to write the generated speech output directly to a WAV file on disk — which is exactly the requirement.

Why the other options are wrong:

AudioOutputConfig(stream) — This routes audio output to an in-memory audio stream object, not a file. It is used when you want to process or play the audio programmatically without saving it to disk.

AudioStreamFormat(wave_stream_format=AudioStreamWaveFormat.PCM) — This class defines the format of an audio stream (e.g., PCM encoding, sample rate). It is used when configuring custom audio streams, not when specifying a file output destination. It is not a valid argument for speechsdk.audio. in this context.

The correct and complete line is:

audio_config = speechsdk.audio.AudioOutputConfig(filename= " output.wav " )

To minimize unintended bias, the training data should be representative of the population for which the model will be used. Microsoft responsible AI guidance identifies fairness as a key principle and notes that discrepancies can occur when systems underperform for specific demographic groups or rarely observed input conditions in training data.

A global insurance dataset may not represent UK drivers. A completely random sample may still underrepresent important groups. Removing protected characteristics alone does not guarantee fairness, because other fields can act as proxies. The best answer is to use a training sample representative of the target UK population.

A recycling machine that uses cameras to identify bottles of the correct shape and reject other items requires analysis of visual input. This is a computer vision workload.

Computer vision is used for image and video analysis tasks such as object detection, classification, shape recognition, and visual inspection.

The other options are incorrect:

A. knowledge mining extracts insights from large collections of documents or data.

B. natural language processing analyzes or generates language.

C. generative AI creates new content, but the requirement is to inspect camera images and identify objects.

For a vision-enabled GPT model deployment, use the Azure OpenAI Responses API to send text and image inputs in a single request. Microsoft’s Foundry SDK documentation states that the endpoint exposes the Responses API and provides access to Azure OpenAI models, including chat completions and multimodal-capable model interactions.

Statement 1: Voice Live returns only transcribed text. = No

Voice Live is not limited to transcription. Microsoft documentation states that the Voice Live API supports real-time bidirectional voice applications, including speech recognition, text-to-speech synthesis, avatar streaming, animation data, and audio processing.

Statement 2: Voice Live requires you to separately implement speech to text and text to speech services. = No

Voice Live provides a single real-time voice API experience rather than requiring separate STT and TTS implementations for the conversational loop. Microsoft describes live AI voice conversations as combining speech capabilities for real-time interaction, and the Voice Live API includes speech recognition and text-to-speech synthesis features.

Statement 3: Voice Live combines speech to text, reasoning, and text to speech into a single conversational experience. = Yes

This is correct. Microsoft’s guidance explains that Azure OpenAI audio/realtime capabilities are for scenarios that combine audio with language understanding, reasoning, or generation in a single model call, and Voice Live supports real-time voice-enabled applications over WebSocket connections.

When calling a Foundry agent programmatically by using the Microsoft Foundry SDK, the application needs the agent’s configuration values, including the agent ID . The Foundry playground provides a Code view that shows sample application code and environment variables needed to call the agent.

The other options are incorrect:

A. copy the tool name used by the agent identifies a tool, not the agent.

B. copy the model family name identifies the model family, not the agent instance.

C. copy the project name identifies the project, not the specific agent ID.

Statement 1: In the new Microsoft Foundry portal, you must fine-tune a model before you can deploy the model. = No

Fine-tuning is optional. Microsoft’s Foundry model deployment documentation describes deploying Foundry Models directly from the model catalog for inference. It does not require fine-tuning first.

Statement 2: In the new Microsoft Foundry portal, you can test a model from the model catalog only after you deploy the model. = No

Microsoft documentation states that some Foundry Tools are available to try via the model catalog without a project , and Foundry playgrounds are used for prototyping and validation before production. Therefore, the statement using “only after you deploy” is too restrictive.

Statement 3: In the new Microsoft Foundry portal, you can deploy a model from the model catalog only after retraining the model. = No

Retraining/fine-tuning is not required before deployment. Microsoft states that after you deploy a Foundry Model, you can interact with it in the Foundry Playground and use it from code, and the deployment workflow starts by selecting a model from the model catalog and choosing Deploy .

Statement

Answer

Generative AI models predict the next token in a sequence based on patterns learned during training.

Yes

A system prompt can be used to influence the behavior and tone of the responses generated by a generative AI model.

Yes

Increasing the Temperature parameter makes the responses generated by generative AI models more deterministic and consistent.

No

Comprehensive and Detailed 150 to 250 words of Explanation From Azure AI Fundamentals/Course Guide/topics:

The first statement is Yes . Generative language models generate text autoregressively by estimating probabilities for the next token based on preceding context and patterns learned during model training. This next-token prediction mechanism is fundamental to how transformer-based generative models produce sequences of text.

The second statement is also Yes . A system prompt provides high-level instructions that establish how the model should behave. It can define role, response style, tone, constraints, and other behavioral expectations. The AI-901 study guide explicitly includes creating effective system and user prompts as part of implementing generative AI applications in Microsoft Foundry.

The third statement is No . Increasing Temperature does the opposite of making output more deterministic. Microsoft documents that higher temperature values increase randomness and output diversity, while lower values make responses more focused and predictable. At a temperature approaching zero, model selection becomes substantially more deterministic because higher-probability tokens are favored more strongly.

The Microsoft responsible AI principle of inclusiveness means AI systems should be designed to empower and engage everyone, including people with different abilities, languages, and accessibility needs.

Therefore, designing the interface to support multiple languages and screen readers is an example of inclusiveness .

Why the other options are incorrect:

A. Provide explanations about how predictions are generated = Transparency

C. Evaluate model outputs across demographic groups to reduce bias = Fairness

D. Encrypt stored customer data and restrict access by using role-based controls = Privacy and security

Statement 1: A summarization model is used to transcribe spoken audio into written text. = No

Summarization creates a shorter version of existing text. Transcribing spoken audio into written text is a speech recognition / speech-to-text capability, not summarization.

Statement 2: A chat completion model can generate vector embeddings for semantic tasks. = No

Chat completion models are used to generate conversational text responses. Vector embeddings for semantic search, similarity, and retrieval tasks are generated by embedding models , not chat completion models.

Statement 3: An embedding model can generate vector representations of the provided input. = Yes

This is correct. Embedding models convert input such as text into vector representations that capture semantic meaning and can be used for tasks such as similarity search, clustering, and retrieval.

The correct answer is Creating captions for a video recording . Speech recognition, also called speech-to-text , converts spoken audio into textual output. Microsoft explicitly defines Azure Speech-to-Text as speech recognition technology that transcribes audio streams or prerecorded audio into text.

Caption generation is a direct application of this capability. Microsoft describes captioning as converting the audio content of a video, film, webcast, or other production into text and displaying that text visually. Therefore, creating captions from the spoken content of a video is a canonical speech-recognition workload.

Creating an audio commentary is instead associated with speech synthesis or text-to-speech because the required output is spoken audio. Identifying key phrases in a video transcript is a text-analysis/NLP task performed after the speech has already been transcribed. A voice-activated security system may involve voice authentication, speaker recognition, or command recognition, but it is not as unambiguously a speech-to-text workload as caption generation.

The AI-901 Study Guide specifically requires candidates to distinguish the features and capabilities of speech recognition and speech synthesis .

The completed sentence is:

Information extraction solutions that detect and read text in scanned documents and images rely on computer vision .

Detecting and reading text in scanned documents and images is typically done with OCR , which is a computer vision capability. Microsoft describes OCR as text recognition/text extraction that extracts printed or handwritten text from images and documents.

Named Entity Recognition (NER) is used to extract dates, quantities, and locations from text.

Named Entity Recognition identifies and categorizes entities in text, such as people, organizations, locations, dates, times, quantities, currencies, and other structured information.

The other options are incorrect:

Key phrase extraction identifies important phrases or main topics.

Language detection identifies the language of the text.

Sentiment Analysis identifies positive, negative, neutral, or mixed sentiment.

Statement 1: An Azure Content Understanding in Foundry Tools analyzer returns results in the HTML format. = No

Azure Content Understanding returns analysis results as structured output, typically JSON , not HTML. Microsoft documentation states that the Content Understanding API returns analysis results in a structured JSON format.

Statement 2: The Content Understanding REST API analyzes content synchronously and returns results in the same response. = No

Content Understanding analysis is asynchronous. The quickstart states that the response header includes an Operation-Location field, which is used to retrieve the results of the asynchronous analysis operation.

Statement 3: Azure Content Understanding in Foundry Tools can transform documents, images, audio, and video files into structured output. = Yes

Microsoft describes Azure AI Content Understanding as processing documents, images, videos, and audio, and transforming them into user-defined output formats.

The completed sentence is:

An AI workload that produces new content based on user input is an example of generative AI .

Generative AI is used to create new content, such as text, images, code, audio, or other outputs, from user prompts or instructions.

The other options are incorrect:

content understanding extracts or interprets information from existing content.

information extraction identifies and extracts structured data from content.

text analysis analyzes existing text for entities, sentiment, key phrases, language, or summaries.