Microsoft AI-900 Azure AI Fundamentals Exam Questions 2025

The chit-chat feature in Azure AI Language's Question Answering (the successor to QnA Maker) is specifically designed to add a conversational layer to a bot. It provides pre-built datasets for common conversational queries, including greetings, small talk, and other user-friendly interactions. By selecting a personality, such as "Professional," you can easily incorporate these responses without having to create them manually. This directly addresses the requirement to make the bot more user-friendly with greetings and other common conversational replies, separate from the core FAQ knowledge base.

A. Increasing the confidence threshold makes the bot stricter about when it provides an answer from the knowledge base; it does not add new conversational content.

B. Active learning is a feature that suggests new question variations to improve the accuracy of the existing knowledge base, not add general conversational greetings.

C. Multi-turn questions are used to create follow-up prompts that guide a user to a specific, refined answer, which is different from adding general chit-chat.

1. Microsoft Learn

Azure AI Language Documentation. "Add chit-chat to a knowledge base." This document states

"Adding chit-chat to your bot makes it more conversational and engaging... When you build your knowledge base

you can add chit-chat from a pre-defined data set." It also details the different personalities available

including "Professional."

2. Microsoft Learn

Azure AI Language Documentation. "Use follow-up prompts to create multiple turns of a conversation." This document explains that multi-turn conversations are used to "refine a question to get to a final answer

" which is distinct from the goal of adding general greetings.

3. Microsoft Learn

Azure AI Language Documentation. "Active learning suggestions." This source describes active learning as a mechanism to "improve the quality of your knowledge base by suggesting alternative questions... based on user submissions." This confirms it is for improving the core Q&A pairs

not for adding conversational filler.

4. Microsoft Learn

Azure AI Language Documentation. "Confidence score." This page explains that the confidence score indicates the confidence in the answer's accuracy. Adjusting the threshold filters results

it does not add new content like greetings.

Azure AI Document Intelligence (formerly Form Recognizer) is a component of the broader Azure AI Services suite. Azure AI Services are a collection of prebuilt, customizable AI models and services available through APIs. The prebuilt receipt model is a specific feature within Azure AI Document Intelligence that allows developers to automatically extract key information from receipts, such as merchant name, transaction date, and totals, without needing to train a custom model. Therefore, Azure AI Services is the correct umbrella service that provides this capability.

A. Azure Al Computer Vision: This is a distinct service within Azure AI Services focused on general image analysis, such as object detection and optical character recognition (OCR), not specialized document extraction.

B. Azure Machine Learning: This is a platform for building, training, and deploying custom machine learning models, not for consuming prebuilt AI models like the receipt model.

D. Azure Al Custom Vision: This service is used to build and train custom models for image classification and object detection, not for analyzing text and structure in documents like receipts.

1. Microsoft Learn. "What is Azure AI Document Intelligence?" Azure AI services documentation. "Azure AI Document Intelligence is an Azure AI service that lets you build automated data processing software using machine learning technology." This establishes Document Intelligence as part of Azure AI Services.

2. Microsoft Learn. "Prebuilt receipt model - Document Intelligence." Azure AI services documentation. This document details the prebuilt receipt model as a feature of the Document Intelligence service.

3. Microsoft Learn. "What are Azure AI services?" Azure AI services documentation. This page lists the portfolio of services

including "Azure AI Document Intelligence (formerly Form Recognizer)" under the main Azure AI Services category.

4. Microsoft Learn. "AI-900: Describe the features of Azure AI Document Intelligence." Study guide for Exam AI-900: Microsoft Azure AI Fundamentals. This official study guide categorizes Document Intelligence as one of the Azure AI Services that provides prebuilt models for document analysis.

Grounding is a prompt engineering technique where a generative AI model is provided with a specific, relevant set of information (the "grounding data") to use as a source of truth for its response. This technique constrains the model to the provided context, improving the factual accuracy and relevance of its output. This is a core principle behind patterns like Retrieval-Augmented Generation (RAG), where the model retrieves information from a knowledge base to "ground" its answer, rather than relying solely on its general training data.

A. providing examples: This technique, known as few-shot learning, shows the model the desired format or style of response through examples, rather than providing a corpus of factual data for a specific query.

B. fine-tuning: This is a process of retraining a model on a specific dataset to adapt its internal parameters (weights). It is not a prompt engineering technique, which works with the model as-is.

D. system messages: A system message sets the overall persona, tone, and behavioral rules for the AI model for an entire conversation, not the specific contextual data for a single response.

1. Microsoft Learn

"Introduction to prompt engineering": Under the "Common prompt engineering techniques" section

it states

"Grounding is the technique of providing the model with information to use in its response. For example

you might provide a text document and prompt the model to use it to answer questions."

2. Microsoft Learn

"Retrieval Augmented Generation (RAG) in Azure AI Studio": In the "What is Retrieval Augmented Generation (RAG)?" section

the documentation explains that RAG improves model performance by "providing it with a source of information as a foundation

" which is the definition of grounding the model with your own data.

3. Microsoft Learn

"AI-900: Describe considerations for generative AI": In the "Describe prompt engineering" unit

the concept of providing context is central. It explains

"The prompt can include contextual information

such as articles or documents that the model can use to respond." This describes the grounding technique.

The process of generating an audio file from a text script is known as speech synthesis, commonly referred to as text-to-speech (TTS). The scenario requires converting written scripts into spoken narration for videos. The Azure AI Speech service provides this exact capability, enabling the creation of natural-sounding synthesized speech from text input. This directly addresses the need to generate an audio file for each video's script.

A. speech recognition: This converts spoken audio into text (speech-to-text), which is the opposite of the required task.

B. language modeling: This is primarily used for understanding and generating human-like text, not for converting text into audio.

D. translation: This converts text from a source language to a target language, not from text to an audio format.

1. Microsoft Learn

"What is text-to-speech?": "Text-to-speech (TTS)

also known as speech synthesis

is the technology that converts text into audible speech. You can use text-to-speech to create synthesized voices for your applications and services." This source directly defines speech synthesis as the correct workload.

2. Microsoft Learn

"AI-900: Describe features of speech workloads on Azure": In the "Speech synthesis (text-to-speech)" section

it states

"Speech synthesis is the capability of software to generate audible speech from text." This aligns perfectly with the question's requirement to generate narration from a script.

3. Microsoft Learn

"What is speech recognition?": "Speech recognition is the process of transcribing spoken audio to text." This reference confirms that speech recognition is the inverse of the required function.

Fine-tuning is the process of customizing a pre-trained foundation model by continuing the training process using your own specific dataset. This involves uploading a set of training examples (often prompt-completion pairs) to Azure OpenAI. The service then uses this data to update the model's weights, adapting its behavior and knowledge to better suit a particular task or domain. This results in a new, custom model that is more specialized than the original base model.

A. completion: A completion is the text generated by the model in response to a prompt; it is the output, not the customization process.

B. grounding: Grounding involves providing the model with specific, timely data at the time of inference (e.g., via RAG), but it does not permanently alter the model's underlying parameters.

D. prompt engineering: This is the technique of carefully crafting input prompts to guide the model's output, which does not involve uploading data to retrain the model.

1. Microsoft Learn

Azure OpenAI Service Documentation. "Customize a model with fine-tuning". This document states

"Fine-tuning improves on few-shot learning by training the model on many more examples than can fit in the prompt. To fine-tune a model

you prepare a dataset of training examples...". This confirms that fine-tuning is the process of training on your own data.

2. Microsoft Learn

AI-900 Learning Path. "Use your own data with Azure OpenAI Service". This module distinguishes between different methods. In the "Understand fine-tuning" section

it clarifies

"Fine-tuning is a technique you can use to customize a model by training it on your own data."

3. Microsoft Learn

Azure OpenAI Service Documentation. "What is fine-tuning for Azure OpenAI Service?". The overview section states

"Fine-tuning is a feature of Azure OpenAI Service that allows you to customize a base model for your specific application... To fine-tune a model

you need to provide a set of training examples...".

Speech recognition, also known as speech-to-text, is the process of transcribing spoken audio into written text. Providing closed captions for videos (B) requires converting the spoken dialogue into text to be displayed on screen. Similarly, creating a transcript of a call or meeting (D) involves converting the spoken words of participants into a written document. Both are core applications of speech recognition technology.

A. An in-car system that reads text messages aloud is an example of text-to-speech (speech synthesis), which converts written text into spoken audio, the opposite of speech recognition.

C. An automated public address system converts pre-written or dynamically generated text into spoken announcements. This is a text-to-speech application, not speech recognition.

1. Microsoft Learn

AI-900: Explore speech capabilities. In the "Transcribe speech input to text" unit

it states

"Speech-to-text is used to transcribe spoken audio into text. Common uses for speech-to-text include creating transcripts of meetings..." This directly supports option D.

2. Microsoft Azure AI Speech service documentation

"What is speech-to-text?". Under the "Key features" section

"Real-time speech-to-text" is described for use cases such as "live captioning

" which directly supports option B.

3. Microsoft Azure AI Speech service documentation

"What is text-to-speech?". This document explains that text-to-speech "enables your applications

tools

or devices to convert text into humanlike synthesized speech." This confirms that scenarios A and C

which involve reading text aloud

are text-to-speech functionalities.

Anomaly detection is a machine learning technique focused on identifying data points, events, or observations that deviate from a dataset's normal behavior. These non-conforming instances are often referred to as anomalies, outliers, or exceptions. A primary application of this technique is to analyze time-series data to detect unexpected spikes, dips, or changes in behavioral patterns. This is crucial for scenarios like fraud detection in financial transactions, monitoring equipment for predictive maintenance, or identifying network security intrusions. The core function is to flag unusual events that warrant further investigation.

A. This describes Natural Language Processing (NLP), which deals with the interaction between computers and human language, not identifying general data anomalies.

B. This describes object detection, a computer vision task that identifies and locates objects within an image, not a general technique for finding unusual data patterns.

D. This describes image classification, a computer vision task that assigns a label to an entire image based on its content, not anomaly detection.

1. Microsoft Learn. (2024). Introduction to anomaly detection. AI-900: Microsoft Azure AI Fundamentals. "Anomaly detection is a machine learning technique that is used to identify unusual patterns or outliers in data." Retrieved from https://learn.microsoft.com/en-us/training/modules/explore-anomaly-detection/1-introduction

2. Microsoft Learn. (2024). What is the Anomaly Detector service?. Azure AI services documentation. "The Anomaly Detector service enables you to monitor and detect abnormalities in your time series data... Anomaly Detector ingests time-series data of any kind and selects the best anomaly detection algorithm for your data to ensure high accuracy." Retrieved from https://learn.microsoft.com/en-us/azure/ai-services/anomaly-detector/overview

3. Microsoft Learn. (2024). Describe machine learning. AI-900: Microsoft Azure AI Fundamentals. This module distinguishes between different machine learning workloads

including anomaly detection

regression

classification

and clustering

clarifying their distinct purposes. Retrieved from https://learn.microsoft.com/en-us/training/modules/describe-machine-learning/

The requirement is to create an original image using Azure OpenAI. The DALL-E model family is specifically designed for this purpose, generating novel images from natural language text descriptions (prompts). Within the Azure AI ecosystem, you would select and deploy a DALL-E model from the model catalog (found in Azure AI Studio) and then provide it with instructions, such as "a peaceful rural scene with a red barn and rolling green hills at sunset," to generate the required image.

A. GPT-3.5 Turbo is a large language model designed for generating and understanding text. It cannot create images.

B. Searching on Microsoft Bing retrieves existing images from the internet. This does not create an original image and may involve copyright issues.

C. GitHub Copilot is an AI assistant for software developers that suggests code and entire functions; it is not a tool for image generation.

1. Microsoft Learn

Azure AI Services Documentation. "Azure OpenAI Service models." This document explicitly categorizes DALL-E models for "Image generation" and GPT-3.5 Turbo models for "Text generation

" clearly distinguishing their functions.

2. Microsoft Learn

AI-900 Courseware. "Module: Explore generative AI on Microsoft Azure

" Unit 4: "Explore image generation." This unit details how generative AI can be used to create images and specifically mentions the DALL-E model as the tool for this task within Azure.

3. Microsoft Learn

Azure AI Services Documentation. "Quickstart: Get started generating images with Azure OpenAI Service." This guide provides a step-by-step tutorial

stating in the introduction

"This quickstart shows you how to use the DALL-E... models to generate images with the Azure OpenAI Service."

The principle of inclusiveness in Microsoft's responsible AI framework is centered on designing AI systems that empower everyone and engage all people. This explicitly includes ensuring that systems are accessible and usable by individuals with diverse abilities, including those with hearing, visual, or other impairments. The scenario's goal of creating a system that serves this wide range of users directly aligns with the core tenet of inclusiveness.

A. fairness: This principle focuses on ensuring AI systems treat all people equitably and mitigate biases related to factors like gender, ethnicity, or other demographics.

C. reliability and safety: This principle ensures that AI systems operate dependably, securely, and consistently as intended, without causing unintended harm.

D. accountability: This principle establishes that the people who design and deploy AI systems are answerable for their operation and impact, emphasizing human oversight.

---

1. Microsoft Learn

"Describe principles of responsible AI" (AI-900 Learning Path).

Section: Inclusiveness. The documentation states

"AI systems should empower everyone and engage people. For people with a hearing

visual

or other impairment

Microsoft AI products and services provide features like real-time captioning

text analytics

and computer vision technology..." This directly supports the correct answer by linking inclusiveness to designing for impairments.

2. Microsoft

"Our approach to responsible AI at Microsoft."

Section: Our principles > Inclusiveness. This official page defines the principle: "Microsoft is committed to creating responsible AI

with inclusiveness as a key principle. This means that AI systems must be designed to be inclusive of all people

regardless of their background

abilities

or identity."

3. Microsoft

"The Responsible AI Standard

v2."

Page 10

Section: 3. Microsoft’s Responsible AI Principles > 3.4 Inclusiveness. The standard details that inclusiveness involves "ensuring that AI systems are designed to be accessible to people with disabilities" and "avoiding the creation or reinforcement of barriers that exclude people from participating in society." This provides a formal definition that matches the question's scenario.

Content filtering is the specific safety system implemented in generative AI services, such as Azure OpenAI, to identify and manage harmful content. This system analyzes both user inputs (prompts) and the AI's generated outputs (responses) in near-real time. It uses a set of classification models to detect and categorize potentially harmful content, including hate speech, violence, and other inappropriate material. When a hateful response is identified, the system can block it from being returned to the user, thereby mitigating the risk of exposing users to harmful content.

A. prompt engineering: This is the practice of designing inputs to guide the AI's output. While it can reduce the likelihood of harmful responses, it is not the mechanism that identifies them.

B. abuse monitoring: This generally refers to tracking service usage patterns (e.g., API call rates) to prevent system misuse, rather than analyzing the specific content of AI responses.

D. fine-tuning: This is a process of further training a model on a custom dataset to specialize its behavior. It is a model development technique, not a real-time safety system for identifying harmful output.

1. Microsoft Learn

"Azure OpenAI Service content filtering." This document states

"Azure OpenAI Service includes a content filtering system that works alongside core models. The system works by running both the prompt and completion through an ensemble of classification models aimed at detecting and preventing the output of harmful content." It explicitly lists Hate as a primary content category that is filtered.

2. Microsoft Learn

"Responsible AI practices for Azure OpenAI models." Under the section "Mitigating harms

" it details that "Content filters are a key component of our responsible AI strategy... to detect harmful content and prevent it from being generated by the service."

3. Microsoft Learn

"Introduction to responsible generative AI." In the "Safety" section

it explains

"Safety systems for generative AI models often include... Content filters to detect and prevent the generation of harmful or inappropriate content." This confirms that content filtering is the designated tool for this purpose.