Microsoft AI-900 - Microsoft Azure AI Fundamentals

According to the Microsoft Azure AI Fundamentals (AI-900) curriculum and Microsoft Learn module “Explore natural language processing”, NLP enables computers to understand, interpret, and analyze human language. One of its key capabilities is sentiment analysis, which detects emotional tone (positive, negative, or neutral) in text.

In this scenario, the chatbot must detect when a customer is upset based on what they type. This directly maps to sentiment analysis, a core NLP function. The Text Analytics service within Azure Cognitive Services provides prebuilt sentiment analysis models that return a sentiment score and classification (e.g., positive, neutral, negative). As per Microsoft Learn, “Natural language processing allows systems to understand sentiment and intent within text and speech to derive meaningful insights.”

Explanation of other options:

A. Anomaly detection identifies unusual patterns in data (e.g., fraud detection), not emotions in text.

B. Semantic segmentation is a computer vision technique used to label pixels in an image.

C. Regression predicts continuous numeric values and is not related to understanding text or emotions.

Therefore, to enable the chatbot to detect when a user is upset based on text input, the correct AI workload is Natural Language Processing (NLP), specifically through Azure Text Analytics sentiment analysis.

Safety system.

According to the Microsoft Learn documentation and the AI-900: Microsoft Azure AI Fundamentals official study guide, the safety system layer in generative AI architecture plays a crucial role in monitoring, filtering, and mitigating harmful or unsafe model outputs. This layer works alongside the model and user experience layers to ensure that generative AI systems—such as those powered by Azure OpenAI—produce responses that are safe, aligned, and responsible.

The safety system layer uses various techniques including content filtering, prompt moderation, and policy enforcement to prevent outputs that could be harmful, biased, misleading, or inappropriate. It evaluates both user inputs (prompts) and model-generated outputs to identify and block unsafe or unethical content. The system might use predefined rules, classifiers, or human feedback signals to decide whether to allow, modify, or stop a response.

In contrast, the other layers serve different purposes:

The model layer contains the core large language or generative model (e.g., GPT or DALL-E) that processes inputs and produces outputs.

The metaprompt and grounding layer ensures the model’s responses are contextually relevant and factually supported, often linking to organizational data sources or system prompts.

The user experience layer defines how users interact with the AI system, including the interface and conversational flow, but does not manage safety enforcement.

Therefore, the layer that uses system inputs and context to mitigate harmful outputs from a generative AI model is the Safety system layer.

This aligns with Microsoft’s responsible AI principles—Fairness, Reliability and Safety, Privacy and Security, Inclusiveness, Transparency, and Accountability—ensuring generative AI operates ethically and safely.

“The Azure OpenAI GPT-3.5 Turbo model can transcribe speech to text.” — NOThis statement is false. The GPT-3.5 Turbo model is a text-based large language model (LLM) designed for natural language understanding and generation, such as answering questions, summarizing text, or writing content. It does not process or transcribe audio input. Speech-to-text capabilities belong to Azure AI Speech Services, specifically the Speech-to-Text API, not Azure OpenAI.

“The Azure OpenAI DALL-E model generates images based on text prompts.” — YESThis statement is true. The DALL-E model, available within Azure OpenAI Service, is a generative AI model that creates original images from natural language descriptions (text prompts). For example, given a prompt like “a futuristic city at sunset,” DALL-E generates a unique, high-quality image representing that concept. This aligns with generative AI workloads in the AI-900 study guide, where DALL-E is specifically mentioned as an image-generation model.

“The Azure OpenAI embeddings model can convert text into numerical vectors based on text similarities.” — YESThis statement is also true. The embeddings model in Azure OpenAI converts text into multi-dimensional numeric vectors that represent semantic meaning. These embeddings enable tasks such as semantic search, recommendations, and text clustering by comparing similarity scores between vectors. Words or phrases with similar meanings have vectors close together in the embedding space.

In summary:

GPT-3.5 Turbo → Text generation (not speech-to-text)

DALL-E → Image generation from text prompts

Embeddings → Convert text into numerical semantic representations

Correct selections: No, Yes, Yes.

According to the Microsoft Azure AI Fundamentals (AI-900) study materials and Microsoft’s Responsible AI guidelines, customers must obtain approval based on their intended usage before accessing and deploying Azure OpenAI Service. This requirement ensures that Microsoft upholds its commitment to Responsible AI principles, which include fairness, reliability and safety, privacy and security, inclusiveness, transparency, and accountability.

The Azure OpenAI Service provides access to powerful language models such as GPT series and Codex, which can generate, summarize, and understand natural language and code. Because of the potential for misuse—such as generating harmful content, misinformation, or unethical automation—Microsoft enforces a use case review and approval process before granting customers access to the service. This process involves submitting an application describing the intended purpose, deployment method, and compliance measures. Only after Microsoft validates that the proposed use aligns with responsible AI practices will access be approved.

This aligns with Microsoft’s documented commitment that “customers are required to submit an application that describes their intended use of the Azure OpenAI Service,” ensuring that all deployments follow ethical and legal standards. This approval step helps maintain transparency and prevent harmful or non-compliant use cases such as deepfake generation, biased automation, or malicious chatbot deployment.

Other options listed in the question are incorrect:

Commit to a minimum level of expenditure – Microsoft does not require financial commitments for ethical approval.

Pay an upfront fee – Payment is handled through normal Azure billing, not a special fee.

Provide credit card details – Not a responsible AI requirement; this is standard for any Azure subscription.

Therefore, the correct and verified answer per Microsoft’s Responsible AI framework and Azure AI-900 study

In the Microsoft Azure AI Fundamentals (AI-900) and Azure Machine Learning (AML) learning paths, deploying a real-time inference pipeline refers to making a trained machine learning model available as a web service that can process incoming data and return predictions instantly. To achieve this, the model must be deployed to an infrastructure capable of handling continuous, low-latency requests with high reliability and scalability.

Microsoft’s official guidance from Azure Machine Learning documentation specifies that:

For testing or development, you can deploy to Azure Container Instances (ACI) because it provides a lightweight, temporary environment suitable for small-scale or non-production workloads.

For production-grade, real-time inference, the deployment should be made to Azure Kubernetes Service (AKS).

AKS provides enterprise-level scalability, load balancing, and high availability, which are critical for serving real-time predictions to multiple consumers simultaneously. It manages containerized applications using Kubernetes orchestration, allowing the model to scale automatically based on traffic demands.

Azure Machine Learning Compute is mainly used for model training and batch inference pipelines, not real-time endpoints. A local web service is typically used only for debugging or offline testing on a developer machine and cannot be shared for external consumption.

Therefore, when deploying a real-time inference pipeline as a service for others to consume, the correct and Microsoft-verified option is Azure Kubernetes Service (AKS). This environment ensures production readiness, secure endpoint management, and scalability for live AI applications, fully aligning with best practices outlined in the Azure Machine Learning designer documentation and AI-900 exam objectives.

https://docs.microsoft.com/en-us/azure/machine-learning/concept-designer#deploy

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and the Microsoft Learn module “Identify features of common machine learning types”, the classification technique is a type of supervised machine learning used to predict which category or class a new observation belongs to, based on patterns learned from labeled training data.

In this scenario, a banking system that predicts whether a loan will be repaid is dealing with a binary outcome—either the loan will be repaid or will not be repaid. These two possible results represent distinct classes, making this problem a classic example of binary classification. During training, the model learns from historical data containing features such as customer income, credit score, loan amount, and repayment history, along with labeled outcomes (repaid or defaulted). After training, it can classify new applications into one of these two categories.

The AI-900 curriculum distinguishes between three key supervised and unsupervised learning approaches:

Classification: Predicts discrete categories (e.g., spam/not spam, fraud/not fraud, will repay/won’t repay).

Regression: Predicts continuous numerical values (e.g., house prices, sales forecast, temperature).

Clustering: Groups data based on similarity without predefined labels (e.g., customer segmentation).

Since the banking problem focuses on predicting a categorical outcome rather than a continuous numeric value, it fits squarely into the classification domain. In Azure Machine Learning, such tasks can be performed using algorithms like Logistic Regression, Decision Trees, or Support Vector Machines (SVMs), all configured for categorical prediction.

Therefore, per Microsoft’s official AI-900 learning objectives, a banking system predicting whether a loan will be repaid represents a classification type of machine learning problem.

According to the Microsoft Azure AI Fundamentals (AI-900) official study guide and Microsoft Learn modules on machine learning concepts, ensuring that the accuracy of a predictive model can be proven requires data partitioning—specifically splitting the available data into training and testing datasets. This is a foundational concept in supervised machine learning.

When you split the data, typically about 70–80% of the dataset is used for training the model, while the remaining 20–30% is used for testing (or validation). The reason behind this approach is to ensure that the model’s performance metrics—such as accuracy, precision, recall, and F1-score—are evaluated on data the model has never seen before. This prevents overfitting and allows you to demonstrate that the model generalizes well to new, unseen data.

In the AI-900 Microsoft Learn content under “Describe the machine learning process”, it is explained that after cleaning and transforming the data, the next essential step is data splitting to “evaluate model performance objectively.” By keeping training and testing data separate, you can prove the reliability and accuracy of the model’s predictions, which is particularly crucial in sensitive domains like clinical or healthcare analytics, where decision transparency and validation are vital.

Option A (Train the model by using the clinical data) is incorrect because you should not train and evaluate on the same data—it would lead to biased results.

Option C (Train the model using automated ML) is incorrect because automated ML is a method for training and tuning, but it doesn’t inherently prove accuracy.

Option D (Validate the model by using the clinical data) is also incorrect if you use the same dataset for validation and training—it would not prove true accuracy.

Therefore, per Microsoft’s official AI-900 study content, the verified correct answer is B. Split the clinical data into two datasets.