Free AI-900 Practice Test Questions and Answers (2026)

The Split Data module in Azure Machine Learning designer is specifically designed to partition a single dataset into two or more disjoint sets of rows. This is the standard and required procedure for creating training and validation/testing datasets. The module allows for various splitting modes, including a simple percentage split, which is commonly used to allocate a majority of the data for training (e.g., 70-80%) and the remainder for validation or testing.

A. Select Columns in Dataset: This module is used for feature selection by including or excluding specific columns (features), not for partitioning rows into training and validation sets.

B. Add Rows: This module concatenates two datasets by appending rows from one to another. It is used for combining datasets, not splitting one.

D. Join Data: This module merges two datasets based on common key columns, similar to a SQL join. It combines data horizontally, not partitions it into subsets.

1. Microsoft Azure Official Documentation. "Split Data module". Azure Machine Learning Documentation. Microsoft Learn. This document explicitly states

"Use the Split Data module to divide a dataset into two distinct sets. This module is useful when you need to separate data into training and testing sets."

2. Microsoft Azure Official Documentation. "Module reference - Azure Machine Learning". Azure Machine Learning Documentation. Microsoft Learn. This page provides an overview of data manipulation modules

confirming the distinct functions of Split Data (partitioning rows)

Select Columns in Dataset (feature selection)

Add Rows (concatenation)

and Join Data (merging based on keys).

This scenario is a direct application of the Reliability and Safety principle. This principle mandates that AI systems must be robust, resilient, and operate dependably, especially when faced with unexpected conditions or "edge cases." Data fields with unusual or missing values represent such unexpected conditions. By being designed to not provide a prediction in this situation, the system avoids generating a potentially inaccurate, nonsensical, or harmful output. This "fail-safe" behavior is a core component of ensuring the system is reliable and safe for its users.

Microsoft. (n.d.). The six principles of responsible AI. Microsoft Learn. Retrieved October 24, 2025, from https://learn.microsoft.com/en-us/azure/machine-learning/concept-responsible-ai

Reference: In the "Reliability and safety" section, it is stated: "AI systems need to be reliable and safe... To be reliable, a system should operate as it was originally designed... This requires rigorous testing and validation to ensure that the system responds safely to edge cases..." Missing or unusual values are a classic "edge case."

Microsoft. (2024, September 12). Responsible AI in practice. Microsoft Learn. Retrieved October 24, 2025, from https://learn.microsoft.com/en-us/training/modules/responsible-ai-practice/2-understand-principles

Reference: In the "Reliability and safety" section, the text notes that systems must be "resilient to failure or unexpected conditions." The module emphasizes that a reliable system must "work as expected," and providing a flawed prediction on missing data would be a failure to work as expected.

Zhang, J., Liu, M., & Liu, X. (2023). Reliability of Machine Learning: A Survey. ACM Computing Surveys, 55(9), Article 189. https://doi.org/10.1145/3555806

Reference: Section 3.1, "Data Quality Issues," explicitly identifies "missing values" and "anomalies" as key challenges to machine learning reliability. The paper discusses (Section 4.1) "failure detection" mechanisms, where a system identifies inputs it cannot process reliably, as a strategy to ensure overall system reliability. Refusing to predict is such a mechanism.

Historically, Azure Machine Learning workspaces were offered in two editions: Basic and Enterprise. The primary differentiator was access to low-code and no-code graphical user interface (GUI) tools. The Enterprise edition was required for using the GUI-based Azure Machine Learning designer and the GUI for automated machine learning (AutoML). The Basic edition was focused on a code-first experience using the SDK and CLI.

While the Enterprise edition has since been retired and all its features are now included in the standard workspace, this question reflects the historical distinction that is often tested. Core functionalities like creating compute instances and datasets were available in both editions.

B. Creating a compute instance is a fundamental compute management task that was available in both Basic and Enterprise editions.

D. Creating a dataset is a core data management capability that was available in both Basic and Enterprise editions.

1. Microsoft Azure Documentation

"What's new in Azure Machine Learning?". In the September 2021 archive

the retirement of the Enterprise edition is announced. It states

"All existing Enterprise edition workspaces will be automatically upgraded to the Basic edition

which now includes all the features of the Enterprise edition at no extra cost." This confirms there was a distinction and that features previously exclusive to Enterprise (like the Designer and AutoML GUI) were moved to the standard tier.

2. Microsoft Learn

"What is Azure Machine Learning designer?". Section: "Get started with designer". This document describes the designer as a "drag-and-drop interface" for building machine learning pipelines without writing code

which was a hallmark feature of the former Enterprise edition.

3. Microsoft Learn

"What is automated machine learning (AutoML)?". Section: "When to use AutoML: classification

regression

& forecasting". This page describes the ability to use the Azure Machine Learning studio UI for a "no-code" experience

which was the capability provided by the Enterprise edition.

Optical Character Recognition (OCR) is the specific technology used to convert images of typed, handwritten, or printed text—such as a scanned newspaper article—into machine-readable digital text. This process allows the historian to 'digitize' the articles, making them searchable, editable, and analyzable as text data.

The other options are incorrect:

• Facial analysis is used to detect and analyze human faces, which is irrelevant to digitizing text.
• Image classification would assign a single label to the entire image (e.g., "newspaper" or "article"), not extract its content.
• Object detection would draw bounding boxes around elements like headlines or photos but would not transcribe the words.

Mori, S., Suen, C. Y., & Yamamoto, K. (1992). Historical review of OCR research and development. Proceedings of the IEEE, 80(7), 1029-1058.

DOI: https://doi.org/10.1109/5.156468

Reference (Section I, Introduction): The paper defines the fundamental goal of OCR as the "problem of converting handwritten or machine-printed text into machine-readable text." This directly aligns with the historian's goal.

Holley, R. (2009). How Good Can It Get? Analysing and Improving OCR Accuracy in Large Scale Historic Newspaper Digitisation Programs. D-Lib Magazine, 15(3/4).

DOI: https://doi.org/10.1045/march2009-holley

Reference (Abstract): This article explicitly discusses the use of "Optical Character Recognition (OCR)" as the core technology for "historic newspaper digitisation programs," which is the exact scenario in the question.

Smith, R. (2007). An Overview of the Tesseract OCR Engine. Proceedings of the Ninth International Conference on Document Analysis and Recognition (ICDAR 2007), Vol. 2, pp. 629-633.

DOI: https://doi.org/10.1109/ICDAR.2007.4376991

Reference (Section II.A, Architecture): "The Tesseract OCR engine... takes a binary image as input... and produces text output." This describes the technical process of converting an image (the scanned newspaper) into text (the digitized version).

Zou, Z., Shi, Z., Guo, Y., & Ye, J. (2019). Object Detection in 20 Years: A Survey. arXiv preprint arXiv:1905.05055.

DOI: https://doi.org/10.48550/arXiv.1905.05055

Reference (Section 1, Introduction): This paper defines object detection as a task to "localize and classify" objects. It distinguishes this from image classification (labeling the whole image) and segmentation. Neither task involves extracting text.

Zhao, W., Chellappa, R., Phillips, P. J., & Rosenfeld, A. (2003). Face recognition: A literature survey. ACM computing surveys (CSUR), 35(4), 399-458.

DOI: https://doi.org/10.1145/954339.954342

Reference (Section 1, Introduction): This survey defines face recognition (a component of facial analysis) as an application for "automatically identifying or verifying a person from a digital image." This clearly differentiates it from the text-based task.

The Dataset output visualization feature in Azure Machine Learning designer is the specific tool used to examine the characteristics of data columns. By right-clicking a dataset output connection and selecting 'Visualize', a user can view detailed summary statistics and visualizations (such as histograms) for each column. These visualizations directly display the distribution of values—including count, unique values, missing values, minimum, maximum, mean, and standard deviation—which is essential for feature engineering, data cleaning, and understanding potential feature columns. Other options, like the Normalize Data or Select Columns modules, perform data manipulation but do not offer distribution exploration.

The correct answer is sourced from official Microsoft Azure documentation regarding data flow and visualization in the Azure Machine Learning studio designer environment.

Official Vendor Documentation (Microsoft Azure):

Microsoft. (n.d.). Visualize data in Azure Machine Learning designer. Microsoft Learn Documentation, Section: "To visualize the output of a module" and "Explore data statistics." (Specific instruction on right-clicking and selecting 'Visualize' to view summary statistics and charts).

Microsoft. (n.d.). Data distribution and statistics in Azure Machine Learning designer. Microsoft Learn Documentation. (Details the type of statistics—e.g., histograms, mean, std dev—that define "distribution of values" displayed by the visualization feature).

University Courseware (Reputable Institutions - Example concept justification):

Hastie, T., Tibshirani, R., & Friedman, J. (2009). The Elements of Statistical Learning: Data Mining, Inference, and Prediction. Springer. Section 3.1: "Exploratory Data Analysis" (Justifies that examining distributions/summary statistics is a prerequisite for feature selection and modeling). ISBN: 978-0387848570. (Page numbers would vary based on edition, but this core concept is fundamental).

The goal is to predict automobile price, which is a continuous numerical value. This is a regression task.

1. Select Columns in Dataset: This module is used first to select only the features (e.g., engine size, horsepower) and the label (price) that are needed for training, removing irrelevant columns from the raw dataset.
2. Split Data: After cleaning, the data must be split into two parts. One part is used to train the model, and the second part is used to score (evaluate) the model.
3. Linear Regression: This is the machine learning algorithm used for this regression task. The Train Model module requires two inputs: the untrained algorithm (Linear Regression) and the training dataset (from the Split Data module).

Microsoft. (n.d.). Regression Modules. Microsoft Docs. Retrieved October 24, 2025, from https://docs.microsoft.com/en-us/azure/machine-learning/algorithm-module-reference/regression

Reference: This official documentation page lists Linear Regression as a module for "predict[ing] a numeric value," which directly applies to predicting automobile prices.

Microsoft. (n.d.). Train Model. Microsoft Docs. Retrieved October 24, 2025, from https://docs.microsoft.com/en-us/azure/machine-learning/algorithm-module-reference/train-model

Reference: The "How to use Train Model" section states it requires two inputs: "An untrained machine learning algorithm" (Linear Regression) and "A training dataset" (from the Split Data module).

Microsoft. (n.d.). Split Data. Microsoft Docs. Retrieved October 24, 2025, from https://docs.microsoft.com/en-us/azure/machine-learning/algorithm-module-reference/split-data

Reference: The module description confirms its purpose: "to divide a dataset into two distinct sets," which is the standard procedure for creating training and testing sets.

Microsoft. (n.d.). Select Columns in Dataset. Microsoft Docs. Retrieved October 24, 2025, from https://docs.microsoft.com/en-us/azure/machine-learning/algorithm-module-reference/select-columns-in-dataset

Reference: The documentation confirms this module is used "to choose a subset of columns to use in downstream operations," justifying its placement after the initial dataset.

The solution correctly maps the four fundamental tasks of facial recognition:

1. Verification: This is a one-to-one (1:1) matching process that confirms if two face images belong to the same person.
2. Similarity: This is a one-to-many (1:N) search that finds other faces in a collection that look similar to a query face, even if they are not the same person.
3. Grouping: This task takes a collection of unknown faces and organizes them into clusters based on visual similarity, grouping faces that likely belong to the same person.
4. Identification: This is a one-to-many (1:N) search that compares a query face against a database of known identities to find a match and determine "who" the person is.

Microsoft. (2024). Overview of the Face API. Azure AI services documentation. Retrieved October 24, 2025.

Location: Section: "Face detection and attributes."

Citation: This official documentation defines the core tasks. It explicitly states: "Face verification... checks the likelihood that two faces belong to the same person... This is a 1:1 matching." It defines identification as a "1:N matching" to find a match from a "person group" (a set of known people). It defines grouping as organizing "an unknown set of faces into several groups based on similarity." Finally, it defines "Find similar" as finding "similar-looking faces... from a face list or large face list."

Wang, M., & Deng, W. (2021). Deep Face Recognition: A Survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(9), 5713-5734.

DOI: https://doi.org/10.1109/TPAMI.2021.3083526

Location: Section 2, "Problem Setting."

Citation: This academic survey paper defines the standard problems in facial recognition. It distinguishes "Face Verification" as a 1:1 problem ("determine whether a pair of images belong to the same identity") and "Face Identification" as a 1:N problem ("search for a query face... from a gallery set"). This supports the distinction between the first and last questions.

Karpathy, A., & Fei-Fei, L. (2017). CS231n: Convolutional Neural Networks for Visual Recognition. Stanford University.

Location: Lecture 11 Notes: "Face Recognition."

Citation: This university courseware clarifies the common tasks. It describes verification (a 1:1 check, e.g., "is this the same person?") and identification (a 1:N search against a known database, e.g., "who is this?"). This aligns with the provided solution's mapping.

Azure Form Recognizer, now part of Azure AI Document Intelligence, is the correct service. It is specifically designed to extract text, layout, and key-value pairs from documents. Crucially, it includes a Personally Identifiable Information (PII) detection feature that can be enabled during analysis. This feature identifies sensitive data such as names, addresses, social security numbers, and financial information within the document's content, allowing for its identification or redaction.

A. Custom Vision is an image analysis service used for building custom image classification and object detection models. It does not analyze text content within documents for sensitive information.

B. Conversational Language Understanding (CLU) is designed to build natural language understanding models for conversational applications, such as chatbots, to identify user intents and entities from utterances.

1. Microsoft Learn

Azure AI Document Intelligence Documentation

"PII detection feature": "The PII detection feature can identify and redact sensitive information in the document analysis output... The PII feature is available with the 2023-07-31 (GA) and later API versions and can be enabled on the Read and Layout models by using the features query parameter."

2. Microsoft Learn

AI-900 Study Guide

"Analyze text": This section describes the capabilities of the Azure Cognitive Service for Language

which includes PII detection. However

Form Recognizer (Document Intelligence) is the specialized service for processing forms and documents and integrates this capability directly for document-centric workflows. The question specifically asks about documents.

3. Microsoft Learn

"What is Custom Vision?": "Azure AI Custom Vision is an image recognition service that lets you build

deploy

and improve your own image identifiers." This confirms its purpose is image recognition

not text analysis in documents.

Content filtering is a specific safety system designed to detect and prevent harmful content from being processed or generated by an AI model. In services like Azure OpenAI, this system analyzes both the input prompt and the output response against predefined categories, including hate speech, violence, and self-harm. If content is flagged as hateful, the system takes a configured action, such as blocking the response from being returned to the user, thereby directly preventing the dissemination of hateful material.

A. abuse monitoring: This is a broader, often reactive process for detecting patterns of misuse over time, not the specific mechanism that actively blocks a single hateful response.

C. fine-tuning: This process adapts a model to a specific dataset to improve its performance on a task, which can reduce but not guarantee the prevention of harmful outputs.

D. prompt engineering: This involves crafting inputs to guide the model, but it is not a foolproof safety mechanism and can be bypassed by adversarial prompts (jailbreaking).

1. Microsoft Learn

"Azure OpenAI Service content filtering

" Introduction. The document states

"The 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 core filtered category.

2. Microsoft Learn

"Responsible AI practices for Azure OpenAI models

" Mitigations section. This section details content filtering as a primary mitigation strategy. It explains

"In addition to the safety system

the Azure OpenAI Service features a content filtering system... The system detects and takes action on specific categories of potentially harmful content in both input prompts and output completions."

3. Microsoft Learn

"Responsible AI standard v2

" Section 2.3

Safety. This document outlines Microsoft's principles. Under the Safety principle

it discusses the need to "prevent models from generating harmful content

" a goal directly achieved through technical implementations like content filtering.

The AI solution is evaluating specific attributes and quality metrics of the face (exposure, noise, occlusion), not simply locating it or identifying the person.

• Facial detection is the process of locating a face in an image, typically by providing bounding box coordinates.
• Facial recognition is the process of identifying or verifying a specific individual.
• Facial analysis (or face attribute analysis) is the process of extracting characteristics from a detected face. The scenario's requirements—evaluating exposure, noise, and occlusion—are explicit examples of face attribute analysis.

Microsoft Azure AI Vision Documentation (Official Vendor Documentation)

Reference: "Face detection and attributes." Azure AI Vision documentation.

Citation: In the "Face attributes" section, the service description explicitly lists the attributes that can be returned from analysis. This includes "Occlusion... Exposure... Noise." This confirms that analyzing these specific photographic qualities is a function of face attribute analysis, which is distinct from mere detection or recognition.

ACM Computing Surveys (Peer-Reviewed Academic Publication)

Reference: Han, H., Wang, Z., Zhang, Z., & Shan, S. (2020). "A Survey on Face Attributes Analysis." ACM Computing Surveys, 53(2), Article 40, pp. 1–36.

Citation (DOI): https://doi.org/10.1145/3371909

Justification: This survey paper defines "Face Attributes Analysis" (FAA) as the task of "predicting a set of attributes (e.g., gender, age, expression, or... 'low quality' attributes) given a facial image." The task described in the question (evaluating noise, exposure) falls directly under this academic definition.

IEEE Transactions on Pattern Analysis and Machine Intelligence (Peer-Reviewed Academic Publication)

Reference: Kumar, N., Belhumeur, P. N., & Nayar, S. K. (2009). "FaceTracer: A Search Engine for Large Collections of Images with Faces." IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(5), 827–840.

Citation (DOI): https://doi.org/10.1109/TPAMI.2009.48

Justification: This foundational paper discusses descriptors for faces, distinguishing between "detection" (finding the face) and "analysis" of attributes, such as pose, illumination (exposure), and occlusion, which are used to describe the state of the face in the image rather than the identity of the person.

The Azure Machine Learning designer is a low-code/no-code tool within the Azure ML workspace. Its primary interface is a visual canvas where users build end-to-end ML pipelines by dragging, dropping, and connecting pre-built modules for tasks like data import, transformation, model training, and scoring.

• Option 3 (automatically selecting an algorithm...) is incorrect as this specifically describes Automated ML (AutoML), a separate feature.
• Option 4 (using a code-first notebook experience) is incorrect as this describes using Jupyter notebooks with the Azure ML SDK, which is the code-based alternative to the visual designer.

Microsoft. (2024, May 22). What is Azure Machine Learning designer? Microsoft Learn.

Section: Introduction, Paragraphs 1-2.

Quote: "Azure Machine Learning designer is a drag-and-drop interface... The designer provides a visual canvas where you can add datasets and modules... You connect the modules to create a pipeline draft..."

URL: https://learn.microsoft.com/en-us/azure/machine-learning/concept-designer

Microsoft. (2024, October 17). What is automated machine learning (AutoML)? Microsoft Learn.

Section: "How AutoML works."

Note: This official documentation confirms that "automatically selecting an algorithm" is the function of AutoML, distinguishing it from the designer.

URL: https://learn.microsoft.com/en-us/azure/machine-learning/concept-automated-ml

Duke University. (n.d.). AI for Everyone: Azure ML Designer. Duke AIPI (Artificial Intelligence for Product Innovation).

Section: "Azure ML Designer."

Quote: "Azure ML Designer... is a visual, drag-and-drop environment... In the Designer, you create an ML 'pipeline' on a visual canvas by dragging and dropping 'modules'..."

URL: https://aipi.pratt.duke.edu/azure-ml-designer

Microsoft. (2024, September 24). What are Jupyter notebooks in Azure Machine Learning? Microsoft Learn.

Section: Introduction.

Note: This documentation defines the "code-first notebook experience" as a separate component (Jupyter notebooks) within the Azure ML workspace, confirming it is distinct from the designer.

URL: https://learn.microsoft.com/en-us/azure/machine-learning/concept-notebooks

The sequence represents the standard, fundamental workflow for creating a machine learning model before it can be deployed.

1. Data Preparation: Raw data must first be collected, cleaned, transformed, and structured (e.g., feature engineering, splitting into training and test sets). A model cannot be trained on unprepared data.
2. Model Training: The prepared training dataset is used to "teach" an algorithm, allowing it to learn patterns and create the model.
3. Model Evaluation: After training, the model's performance must be validated using a separate test dataset (which it has not seen before) to measure its accuracy and effectiveness. A model is only deployed if its evaluation results meet the required business objectives.
4. The other options, model retraining and data encryption, are incorrect for this sequence. Model retraining is a post-deployment process performed to update a model when its performance degrades over time (model drift). Data encryption is a security process applied to data at rest or in transit, not a sequential step in the model creation pipeline itself.

Source: Microsoft (n.d.). What is Azure Machine Learning designer? Microsoft Docs.

Reference: In the section titled "Build a pipeline," the documentation describes the visual workflow: "Pipelines start with your data... you can connect these datasets to other modules which prepare data, train models, and evaluate models." This officially confirms the sequential order of these three core processes.

Source: Amershi, S., Begel, A., Bird, C., et al. (2019). Software Engineering for Machine Learning: A Case Study. 2019 IEEE/ACM 41st International Conference on Software Engineering: Software Engineering in Practice (ICSE-SEIP).

Reference: Page 292 (Section III.A, "AI Development Workflow"). The paper describes the typical ML workflow stages identified through case studies, stating: "...the AI development workflow includes data preparation (e.g., collect, clean, label), model training (e.g., feature engineering, training), and model evaluation..."

DOI: https://doi.org/10.1109/ICSE-SEIP.2019.00018

Source: Microsoft (n.d.). The machine learning lifecycle. Microsoft Docs.

Reference: The document explicitly outlines the stages of the machine learning lifecycle. The core model creation process is described in this order: 3. Data preparation, 4. Model training, and 5. Model evaluation, which all occur before stage 6. Model deployment. This clearly separates the pre-deployment steps from post-deployment activities like "Model monitoring" (which leads to retraining).

In Azure Machine Learning, registering a model is the standard process for storing and versioning a trained model within your workspace's model registry. When you register a model with a specific name, it is saved as version 1. If you subsequently train a new version of the model and register it using the same name, Azure Machine Learning automatically increments the version number (e.g., to version 2). This functionality is central to MLOps practices, as it allows you to track, organize, and manage the lifecycle of all your trained models, ensuring reproducibility and enabling you to select a specific version for deployment.

A. Provision an inference duster.

An inference cluster is a compute target used for deploying a model to perform real-time predictions, not for tracking or versioning the model asset itself.

B. Explain the model.

Explaining a model is part of Responsible AI and involves generating interpretations for its predictions. This process does not track or version the model.

D. Register the training data.

Registering data creates a versioned dataset in the workspace. While this tracks the data used for training, it does not track the resulting trained model.

---

1. Microsoft Learn Documentation

"Manage models in Azure Machine Learning": Under the section "Register and track models

" it states

"Registering a model allows you to store and version your models in the Azure cloud

in your workspace. The model registry makes it easy to organize and keep track of your trained models."

2. Microsoft Learn Documentation

"Train and register a model with the CLI (v2)": In the "Register the model" section

the documentation explains

"Now that you've trained the model

you can register it in your workspace... Each time you register a model with the same name

the registry assumes it's a new version."

3. Microsoft Learn Documentation

"What is MLOps?": In the "Model registration" section of the MLOps lifecycle diagram

it is defined as: "Model registration is the act of storing and versioning a model in your Azure Machine Learning workspace." This confirms registration is the key step for versioning.

Automated machine learning (automated ML or AutoML) in Azure Machine Learning is the specific capability designed to automate the iterative and time-consuming tasks of model development. It allows users to build, train, and deploy predictive models with high efficiency and productivity by automatically trying various algorithms and hyperparameter combinations. This process can be initiated through a user-friendly interface in the Azure Machine Learning studio, requiring minimal to no coding, which directly aligns with the requirement to "quickly build and deploy a predictive model without extensive coding."

A. ML pipelines are used to create and manage reproducible workflows for machine learning tasks, which typically requires coding and is focused on productionization and automation, not rapid initial model creation.

B. Copilot is a general term for an AI assistant that helps write code. While it can speed up coding, it does not eliminate the need for it in building a model.

C. DALL-E is a generative AI model within the Azure OpenAI Service used for creating images from text prompts, not for building predictive models from data.

1. Microsoft Learn

AI-900 Study Guide

"Describe automated machine learning": "Automated machine learning (often abbreviated to automated ML or AutoML) is a key capability in Azure Machine Learning that enables non-experts to create an effective machine learning model... It helps data scientists and developers save time by automating the iterative tasks..."

2. Microsoft Learn

"What is automated machine learning (AutoML)?": "Automated machine learning... is the process of automating the time-consuming

iterative tasks of machine learning model development. It allows data scientists

analysts

and developers to build ML models with high scale

efficiency

and productivity all while sustaining model quality."

3. Microsoft Learn

"What are Azure Machine Learning pipelines?": "An Azure Machine Learning pipeline is an independently executable workflow of a complete machine learning task... Pipelines are defined by using a Python SDK

or visually in Azure Machine Learning studio designer." This highlights that pipelines are for creating workflows

not the model itself in a no-code fashion.

4. Microsoft Learn

"What is Azure OpenAI Service?": Under the "Model summary" section

it lists "DALL-E models" as a "series of models that can generate original images from natural language." This confirms its purpose is image generation.

The chit-chat feature in QnA Maker (now part of Azure AI Language's Custom question answering) is specifically designed to make a bot more conversational and engaging. It allows you to add a pre-populated set of conversational responses for common small talk, such as greetings, goodbyes, and other pleasantries. You can select a specific personality for the bot, including "Professional," which directly addresses the user's need to add professional greetings and make the bot more user-friendly without having to manually create these question-and-answer pairs.

A. Increase the confidence threshold of responses: This adjusts how certain the service must be to return an answer. It filters out less relevant answers but does not add new conversational content like greetings.

B. Enable active learning: This feature suggests new question variations to add to your knowledge base based on user queries, improving the accuracy of existing answers, not adding general small talk.

C. Create multi-turn questions: This feature helps guide users to a specific answer through follow-up prompts. It is used for disambiguation within a topic, not for general, out-of-context greetings.

1. Microsoft Learn

"Add chit-chat to a project": "Adding chit-chat to your bot makes it more conversational and engaging. The chit-chat feature in custom question answering allows you to easily add a pre-populated set of the top chit-chat

to your project... When you create your project

you can add a chit-chat data source. You choose the personality you want for your chit-chat. The personalities available are: Professional

Friendly

Witty

Caring

and Enthusiastic."

2. Microsoft Learn

"Confidence score": "The confidence score indicates the confidence that the answer is the correct match for a given user query... By default the threshold is set to 0

so all possible answers are returned." This confirms its role as a filter

not a content source.

3. Microsoft Learn

"Active learning suggestions": "Active learning gives you suggestions for new question and answer pairs to add to your project based on user queries that are hitting your published project." This clarifies its purpose is to improve the existing knowledge base.

4. Microsoft Learn

"Use follow-up prompts to create multiple turns of a conversation": "Custom question answering lets you add follow-up prompts to question and answer pairs. This feature lets your client application provide a determined conversation flow with the user to reveal the correct answer." This shows its function is for guided conversations

not general greetings.

Conversational AI is defined by its ability to engage in a bidirectional dialogue, requiring components like Natural Language Understanding (NLU) to interpret intent and Natural Language Generation (NLG) to formulate a contextual response.

1. A webchat bot (Yes) is a prime example, as its core function is to understand user input and generate a relevant reply, thus maintaining a conversation.
2. A smart device (Yes) acting as a virtual assistant also fits this definition. It must understand the spoken query ("What will the weather be?"), process the intent, retrieve information, and generate a natural language answer.
3. Automatic captioning (No) is an application of Automatic Speech Recognition (ASR). ASR is a component used by conversational AI (to transcribe speech to text) but is not conversational AI by itself. It is a one-way transcription process and does not involve understanding intent or generating a dialogue-based response.

López, G., et al. (2020). "A review on conversational artificial intelligence." IEEE Access, vol. 8, pp. 137976-137994.DOI: https://doi.org/10.1109/ACCESS.2020.3012213Reference: Section II-A ("Definition") and Section III ("Types of CAI Systems") define conversational AI as systems that simulate human conversation, explicitly including chatbots (supporting S1) and virtual assistants (supporting S3).Ram, A., et al. (2018). "Conversational AI: The Science Behind the Technology." ACM Queue, vol. 16, no. 5.DOI: https://doi.org/10.1145/3283250.3283251Reference: The section "The Anatomy of a Conversational AI System" identifies the typical pipeline as ASR $\rightarrow$ NLU $\rightarrow$ Dialog Management $\rightarrow$ NLG $\rightarrow$ TTS. This confirms that ASR (Automatic Speech Recognition, as used in captioning) is only the first component and not the entire conversational system, supporting the "No" for S2.Massachusetts Institute of Technology (MIT). (2019). "Lecture 9: Chatbots." MIT 6.S198: Deep Learning.Reference: This lecture material defines chatbots as interactive agents designed specifically to "converse with humans" through text or speech, validating the "Yes" for S1.Badr, M., et al. (2022). "A Survey on Automatic Speech Recognition in Conversational AI." 2022 IEEE International Conference on Big Data (Big Data).DOI: https://doi.org/10.1109/BigData55660.2022.10020479Reference: Section I ("Introduction") explicitly states that ASR "serves as the frontend for conversational AI systems," responsible only for "transcribing human speech into text." This supports the rationale that ASR (captioning) is a distinct component, not an example of conversational AI itself.

App1 requires the extraction of text (brand names) from visual data (images/photos). Optical character recognition (OCR) is the precise AI technique designed for recognizing and extracting text from images, making it the most accurate choice over the broader Image Analysis. App2 describes a system for touchless access control, which universally relies on biometric identification, specifically Face recognition and verification, for authentication to grant entry. The Azure AI Vision Face service provides the necessary capabilities for identifying individuals to enable this functionality.

Microsoft Official Documentation (Azure AI Services - Vision):

Topic: Azure AI Vision Overview and Capabilities (OCR).

Reference: Official vendor documentation on the Read API or OCR, detailing its ability to "extract text from images and documents," which includes text like brand names. (Specific section on Optical Character Recognition).

IEEE Xplore Digital Library:

Topic: Face Recognition in Access Control Systems.

Reference: An academic paper discussing the application of face recognition algorithms (specifically detection, verification, and identification) for secure access systems. (e.g., IEEE Access, Section on Biometric Systems).

MIT OpenCourseWare (Courseware on Computer Vision):

Topic: Fundamentals of Optical Character Recognition.

Reference: Course materials defining OCR as the method for converting "scanned images of text into machine-encoded text," directly applicable to extracting text like brand names from photos. (e.g., Lecture notes on Document Analysis).

Cambridge University Press (Academic Publication on Biometrics):

Topic: Biometric Authentication for Physical Access.

Reference: Textbook chapter explaining that face recognition is a common, non-contact biometric method used for access control and security applications. (e.g., Chapter on Non-Contact Biometrics, pages 70-75).

Azure Form Recognizer is an AI service designed to extract text, key-value pairs, and tables from documents. It includes prebuilt models specifically trained for common document types like invoices and receipts. The service can identify and extract specific fields, such as an invoice number from an invoice or the retailer's name from a sales receipt. These capabilities make it the ideal solution for automating data entry from these types of forms.

B. Translating a form is a language translation task. The appropriate service for this is the Azure AI Translator service, not Form Recognizer, which focuses on data extraction.

C. Finding a product image in a catalog is an object detection or image analysis task. This would be handled by Azure AI Computer Vision or Custom Vision services.

1. Microsoft Learn

"What is Azure Form Recognizer?": This document states

"Azure Form Recognizer is a cloud-based Azure Applied AI Service that uses machine-learning models to extract key-value pairs

text

and tables from your documents." It explicitly lists invoices and receipts as primary use cases.

2. Microsoft Learn

"Form Recognizer prebuilt invoice model": Under the "Field extraction" section

this document lists InvoiceId as a standard field that the prebuilt invoice model extracts

directly supporting option A.

3. Microsoft Learn

"Form Recognizer prebuilt receipt model": The "Field extraction" section for the receipt model lists MerchantName as a key field it extracts from sales receipts

which corresponds to identifying the retailer in option D.

4. Microsoft Learn

"AI-900: Describe features of natural language processing (NLP) workloads": This module distinguishes between Azure AI services. It positions the Translator service for "translating text from one language to another" (refuting option B) and Form Recognizer for "extracting information from forms

receipts

and invoices."

This sequence represents the fundamental workflow for supervised machine learning, which is visually implemented in the Azure Machine Learning designer.

1. First, data must be ingested and prepared (e.g., cleaned, transformed).
2. Next, this dataset is split into a training set (to teach the model) and a validation/test set (to test it).
3. The model is then trained using only the training data.
4. Finally, the trained model's performance is assessed by evaluating it against the unseen validation dataset. Evaluating against the "original dataset" would be incorrect as it includes the training data, leading to an over-optimistic and invalid performance metric (overfitting).

Microsoft Press Store (for Exam DP-100): In the guide "Explore data and train models," the process for creating a training pipeline in the designer is detailed. It specifies the sequence:

Use modules to "clean and transform your data" (Prepare).

"Employ the Split Data module to divide your dataset into training and validation sets."

"Connect your preprocessed and split datasets to the Train Model module."

"Examine the output of the Evaluate Model module to assess the performance."

(Source: Microsoft Press Store, "Explore data and train models," Article 2, Skill 2.2: "Create models by using the Azure Machine Learning Designer," sections "Create a training pipeline" and "Here is a more detailed set of instructions...")

Microsoft Learn (Azure ML Documentation): The tutorial "Tutorial: Train a no-code regression model in the designer" demonstrates this logical flow. A dataset is added, prepared (e.g., Select Columns in Dataset, Clean Missing Data), and then split using the Split Data component. The outputs of the split data component are then fed into the Train Model component, which is subsequently connected to a Score Model and Evaluate Model component.

(Source: Microsoft Learn, "Tutorial: Train a no-code regression model in Azure Machine Learning designer," sections "Prepare data," "Split the data," "Train the model," and "Test, score, and evaluate the model.")

Microsoft Learn (Component Reference): The official documentation for the Train Model component states it has two inputs: an untrained algorithm and the "training dataset." It explicitly notes, "After training is completed, use the trained model with one of the scoring components, to make predictions on new data," which is the function of the Evaluate Model component.

(Source: Microsoft Learn, "Train Model: Component Reference," sections "How the training process works" and "How to use Train Model.")

The Azure AI Language service is the correct choice for determining a user's intent in a chatbot. Specifically, its Conversational Language Understanding (CLU) feature is designed to build custom natural language models. These models analyze conversational text to predict the user's intention (intent) and extract important information (entities). This capability is fundamental for a chatbot to understand what a user wants and to trigger the appropriate action or response. CLU is the designated service in Azure for building the core language understanding component of a conversational AI solution.

A. Translator: This service is used for translating text from one language to another; it does not analyze or determine the underlying intent of the text.

B. Azure Cognitive Search: This is an information retrieval service for indexing and querying large datasets. It finds relevant documents, but does not identify specific, pre-defined user intents for a conversational flow.

C. Speech: This service converts spoken audio to text (speech-to-text) and text to audible speech (text-to-speech). It processes the modality of communication, not the semantic meaning or intent.

---

1. Microsoft Learn

Azure AI Language Documentation. "What is conversational language understanding (CLU)?". This document states

"Conversational language understanding (CLU) is a cloud-based conversational AI service that applies custom machine-learning intelligence to a user's conversational

natural language text to predict overall meaning

and pull out relevant

detailed information." This directly confirms its purpose for identifying intent.

Source: Microsoft Azure Official Documentation

https://learn.microsoft.com/en-us/azure/ai-services/language-service/conversational-language-understanding/overview

Section: "What is conversational language understanding (CLU)?"

Paragraph 1.

2. Microsoft Learn

AI-900 Exam Study Guide. The official study guide for the AI-900 exam lists "Understand conversational language understanding" as a key skill measured. It places this capability within the "Azure AI Language" service

reinforcing that this is the correct service for building conversational solutions like chatbots.

Source: Microsoft Learn

https://learn.microsoft.com/en-us/credentials/certifications/exams/ai-900/

Section: "Skills measured"

Subsection: "Describe features of natural language processing (NLP) workloads on Azure".

3. Microsoft Learn

Azure AI Speech Service Documentation. "What is the Speech service?". This document clarifies the service's function: "The Speech service provides speech-to-text and text-to-speech capabilities with a Speech resource." This confirms it does not handle intent recognition.

Source: Microsoft Azure Official Documentation

https://learn.microsoft.com/en-us/azure/ai-services/speech-service/overview

Section: "What is the Speech service?"

Paragraph 1.