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

Azure Content Understanding uses transcription to convert audio content, such as voicemail recordings, into text. Microsoft’s Azure Content Understanding audio documentation states that transcription converts conversational audio into searchable and analyzable text-based transcripts.

Option A. Voice Live is not the feature used by Content Understanding to convert voicemail recordings into text.

Option B. key phrase extraction identifies important phrases after text is available; it is not the audio-to-text conversion step.

Option D. optical character recognition (OCR) is for extracting text from images or documents, not audio.

Therefore, the correct answer is C. transcription.

Statement 1: In the Foundry playground, you can upload a local image and include text in the same message when prompting a multimodal model. = Yes

Microsoft documentation for vision-enabled chat models states that in the chat session pane, you can select the attachment button, upload an image, add a text prompt such as “Describe this image,” and then submit it.

Statement 2: When using the OpenAI Responses API and a vision-enabled model, images can be provided only as base64-encoded image data. = No

Images are not limited to base64 data. The Responses API supports image input through an image URL, and Azure OpenAI documentation also describes examples using base64 encoded image data.

Statement 3: Prompts that include images require deploying a text-only model because multimodal capabilities are handled by the application layer. = No

Image prompts require a vision-enabled multimodal model, not a text-only model. Azure OpenAI vision-capable models accept image and text inputs directly.

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.

client = OpenAI(

api_key= " ... " ,

base_url= " https://resource1.openai.azure.com/openai/v1/ " ,

)

response = client.responses.create(

model= " my-mini-gpt " ,

)

For Azure OpenAI in Microsoft Foundry, the base_url uses the Azure OpenAI resource name in the endpoint format:

https:// < resource-name > .openai.azure.com/openai/v1/

In the question, the Azure OpenAI resource is named Resource1, so the first blank must be resource1. Microsoft documentation for Azure OpenAI v1 endpoints confirms that the endpoint must use the ...openai.azure.com/openai/v1/ path.

For the model parameter, Azure OpenAI requires the deployment name, not the underlying model name. Microsoft states that Azure OpenAI always requires the deployment name when calling APIs, even when the parameter is named model.

The deployed model is gpt-4.1-mini, but the deployment name is my-mini-gpt. Therefore, the second blank must be:

model= " my-mini-gpt "

So the correct selections are:

base_url blank = resource1

model blank = my-mini-gpt

To compare the costs of large language models in Microsoft Foundry portal, use the Model leaderboard.

Microsoft documentation states that the model leaderboard helps compare models across quality, safety, estimated cost, and throughput. It also supports trade-off charts and side-by-side model comparison for features, performance, and estimated cost.

Why the other options are incorrect:

A. Evaluator catalog is for selecting evaluators to measure model or application outputs, not comparing LLM costs.

C. Compliance relates to governance and compliance, not model cost comparison.

D. Tools provides Foundry tools, not benchmarked cost comparison across models.

The correct answer is classify email messages as work-related or personal. Natural language processing (NLP) is concerned with analyzing, interpreting, and processing human language, including unstructured text contained in documents, messages, reviews, and emails. Text classification is therefore a direct NLP workload: an AI model analyzes the linguistic content of an email and assigns it to a predefined category such as work-related or personal.

Microsoft ' s current AI-901 study guide explicitly requires candidates to identify scenarios for common AI workloads, including text analysis, and to understand common text-analysis techniques. Microsoft Azure Language in Foundry Tools similarly supports capabilities for extracting information from text and classifying text, demonstrating that categorization based on textual content is an NLP function.

Predicting future car rentals is primarily a regression/forecasting problem because the output is a numerical quantity. Predicting whether a website visitor will transact is typically a binary classification machine-learning problem based on behavioral or transactional features, not necessarily language. Stopping industrial machinery when a temperature threshold is exceeded is a rule-based monitoring/control scenario.

Generating marketing email drafts from a short product description and target audience is a content generation task. This is a generative AI workload because the system creates new text based on the user’s prompt.

B. computer vision is for interpreting images or video.

C. text classification categorizes existing text, but does not draft new marketing emails.

D. speech recognition converts spoken audio into text.

Therefore, the correct answer is A. generative AI.

The Microsoft responsible AI principle of transparency requires that AI systems be explainable to users.

Transparency means users should understand when they are interacting with AI, what the system can and cannot do, and how outputs or predictions are generated at an appropriate level.

The other options map to different responsible AI principles:

protect sensitive user data = Privacy and security

reduce bias in decisions = Fairness

require human oversight = Accountability

The system must not discriminate based on gender, race, or age. = Fairness

This requirement maps to fairness because fairness means AI systems should avoid unfair bias and treat people equitably across demographic groups.

Personal data must be visible only to approved users. = Privacy and security

This requirement maps to privacy and security because it focuses on protecting personal data and restricting access to authorized users only.

Automated decision-making processes must be recorded so that approved users can identify why a decision was made. = Transparency

This requirement maps to transparency because recording decision-making processes helps users understand and explain how or why an AI-supported decision was made.

Azure Content Understanding analysis operations are asynchronous long-running operations. Microsoft’s Python SDK documentation states that analysis starts with a begin operation, then the application polls for results until analysis completes.

Therefore, after submitting the PDF for analysis, the application should poll the operation results until the asynchronous analysis job succeeds.

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 send chat prompts to a deployed model, the application needs authentication credentials, the endpoint URL, and the model deployment name. Microsoft’s Foundry SDK guidance shows authentication with credentials, an endpoint, and the model= " model_deployment_name " value when calling a deployed model.

To minimize cost and keep generated responses within a defined length, configure Max Completion Tokens.

Microsoft’s Azure OpenAI / Foundry API reference defines max_completion_tokens as an upper bound for the number of tokens that can be generated for a completion. Because generated tokens contribute to usage and response length, limiting completion tokens helps control both output length and cost.

Temperature and Top P control randomness or sampling behavior, not maximum response length. Model version settings do not directly define the generated response length.

images.generate

The canonical Python call is:

response = client.images.generate(

model= " gpt-image-1 " ,

prompt=prompt

)

The correct selection is images.generate . Microsoft’s current Foundry documentation shows that image generation with a deployed GPT Image model uses the OpenAI Python SDK ' s Images API , specifically client.images.generate(). Microsoft provides an example using model= " gpt-image-1 " together with a text prompt, which directly corresponds to this scenario.

chat.completions.create is intended for chat-completion interactions rather than the standalone Images API. images.create is not the documented Python SDK method for generating an image. responses.create belongs to the Responses API , which can also orchestrate image generation through an image-generation tool, but that is a different API pattern from the explicitly requested OpenAI Images API . Microsoft distinguishes the standalone image-generation API from image generation performed through the Responses API.

One important detail in the supplied exhibit is that it visually shows client.responses. before the dropdown. Taken literally, client.responses.images.generate() would not be the documented syntax. For this question, the intended answer is clearly images.generate , yielding the documented form client.images.generate(...) .

Study Guide Reference: Microsoft Azure AI Fundamentals → Implement AI solutions by using Microsoft Foundry → Generative AI and image-generation models.

To automate masking of names and phone numbers in text data, use Personally Identifiable Information (PII) detection in Azure Language.

PII detection identifies sensitive personal information such as names, phone numbers, email addresses, addresses, and other personal identifiers. It can be used to detect and redact or mask that information from text.

The other options are incorrect:

A. custom text classification categorizes documents or text into labels.

C. entity linking identifies entities and links them to knowledge base entries.

D. custom named entity recognition (NER) can extract custom entities, but for built-in masking of names and phone numbers, PII detection is the correct feature.

Statement 1: Fairness can be achieved by focusing solely on improving the overall accuracy of an AI model. = No

Overall accuracy alone does not guarantee fairness. A model can have high overall accuracy while still producing biased or lower-quality outcomes for specific demographic groups. Fairness requires evaluating and reducing differences in model behavior across groups.

Statement 2: Evaluating AI system outputs to identify and reduce bias across demographic groups supports fairness. = Yes

This statement is correct. The Microsoft responsible AI principle of fairness means AI systems should treat people fairly and avoid unfair bias. Evaluating outputs across demographic groups helps identify and reduce bias.

Statement 3: Ensuring fairness for an AI system means that all users always receive the same output from the system. = No

Fairness does not mean identical outputs for all users. It means the system should avoid unfair bias and provide equitable treatment. Different users may receive different outputs when differences are relevant and appropriate.

The requirement is to extract structured fields, including nested values, from invoices by using a defined schema. In Azure Content Understanding, an analyzer is the processing unit that defines how content is analyzed, what information is extracted, and how the output is structured, including JSON fields.

Microsoft’s Content Understanding document solutions documentation states that Content Understanding uses customizable analyzers to extract essential information, fields, and relationships from documents and forms. Microsoft’s quickstart also shows invoice processing with the prebuilt-invoice analyzer to extract structured data from an invoice document.

Why the other options are incorrect:

A. transcription workflow in Azure Speech is for converting audio to text, not invoice field extraction.

B. OCR-only document processing can extract text but does not meet the requirement for structured fields and nested values by schema.

D. Azure AI Search is for indexing and querying content, not defining invoice extraction schemas.

Therefore, the correct answer is C. an analyzer in Azure Content Understanding in Foundry Tools.

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.

The completed sentence is:

Ensuring that human reviewers oversee AI-generated decisions and remain responsible for the final output is an example of the Microsoft responsible AI principle of accountability.

Accountability means people and organizations remain responsible for AI systems and their effects. Human review and responsibility for final decisions are examples of accountability.

The other options are incorrect:

fairness focuses on avoiding bias and treating people fairly.

privacy and security focuses on protecting data and restricting access.

transparency focuses on explaining AI use, capabilities, and limitations.

For a vision-enabled model, Microsoft documentation shows that multimodal prompts can include both text and image content in the same user message content array. Azure OpenAI vision models accept multimodal image-and-text input and return natural language responses.

The correct request structure is a user message with a content array containing both items, for example:

{

" role " : " user " ,

" content " : [

{ " type " : " input_text " , " text " : " What is in this image? " },

{ " type " : " input_image " , " image_url " : image_url}

]

}

This avoids sending separate requests and is therefore the best choice for quickest response time.

A and C are incorrect because they send the text and image separately.

B is incorrect because the image/question should be sent as a user message, while system messages are used for instructions and behavior.

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 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.

Azure Text Analytics client library

Comprehensive and Detailed Explanation with all Azure AI documents: =

The completed sentence is:

To develop an application that analyzes text by using Azure Language in Foundry Tools, use the Azure Text Analytics client library package.

Azure AI Language text analysis features, such as sentiment analysis, named entity recognition, key phrase extraction, language detection, and text summarization, are accessed in code by using the Azure Text Analytics client library.

The other options are incorrect:

Azure AI Projects client library is used for Foundry project and agent-related workflows, not specifically Azure Language text analysis.

Azure OpenAI client library is used to call OpenAI model deployments, not Azure Language text analytics APIs.

Azure Speech SDK is used for speech-to-text and text-to-speech workloads, not text analysis.

Speech

The correct answer is speech . The scenario specifically describes converting spoken customer requests into written text , which is a speech-to-text , or speech recognition, capability. Microsoft classifies speech as one of the common AI workloads in the current Azure AI Fundamentals study guide. The guide explicitly requires candidates to identify scenarios involving speech and to understand the capabilities of speech recognition and speech synthesis .

Microsoft Azure Speech in Foundry Tools provides speech-to-text functionality that converts audio streams into textual transcriptions. It supports real-time transcription, fast transcription, and batch transcription depending on the workload.

Key phrase extraction analyzes existing text to identify important words or concepts; it does not perform audio transcription. Text analysis covers operations such as sentiment analysis, entity recognition, summarization, and key phrase extraction after text already exists. Natural Language Understanding (NLU) interprets the meaning or intent of language, but the conversion from audio into written words itself is specifically a speech recognition workload .

Study Guide Reference: Identify AI concepts and capabilities → Identify AI workloads → Identify features and capabilities of speech recognition and speech synthesis.

The chatbot must interpret user intent and route actions such as accepting orders, retrieving support documents, and retrieving order status. This is language understanding, because the system must understand what the user wants to do from natural language input.

Microsoft’s Foundry Agent Service documentation states that tool_choice provides deterministic control over tool calling:

auto means the model decides whether to call tools.

required means the model must call one or more tools.

none means the model does not call tools.

Therefore:

A. auto = Incorrect, because the model may or may not call the Azure function.

B. none = Incorrect, because this prevents tool/function calls.

C. required = Correct, because it forces the agent to call a tool.

The Azure OpenAI function-calling documentation also confirms that tool_choice= " auto " lets the model decide whether to call a function, while tool_choice= " none " forces a user-facing response without a tool call.

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.

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 ensure a vision-enabled model produces a relevant and useful response, the prompt should include a clear description of the task. For example, you should tell the model what to do with the image, such as describe it, identify objects, compare images, extract visible text, or answer a specific question about the visual content.

The other options are incorrect:

B. a dataset ID for the image is not needed in the prompt.

C. the version number of the model is deployment/configuration information, not prompt content.

D. the deployment name of the model is used in the API call to route the request, not inside the prompt.

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.