AI-102 — Azure AI Engineer Associate

Full question text, answer options, and explanations for this practice set — a spoiler-free alternative is the interactive quiz above for scored, shuffled practice.

1. 1. Your company wants to build an AI solution that generates product descriptions using a large language model. You need to choose the correct Azure AI service. Which service should you select?

   • A. Azure AI Vision
   • B. Azure OpenAI in Foundry Models(correct)
   • C. Azure AI Speech
   • D. Azure AI Document Intelligence

   Explanation: Azure OpenAI in Foundry Models provides access to large language models such as GPT-4 for text generation tasks like writing product descriptions. Azure AI Vision handles image analysis (A), Azure AI Speech handles voice/audio processing (C), and Azure AI Document Intelligence extracts structured data from documents (D) — none of these generate free-form product text.

2. 2. A development team needs to authenticate API calls to an Azure AI multi-service resource. The security policy requires that credentials not be stored in application code or configuration files. Which authentication method best satisfies this requirement?

   • A. Store the subscription key in an environment variable on the server
   • B. Use a managed identity assigned to the compute resource and grant it the Cognitive Services User role(correct)
   • C. Embed the API key directly in the request header in source code
   • D. Rotate the API key every 30 days and redeploy the application

   Explanation: Managed identities eliminate the need to store credentials anywhere in code or configuration — the Azure platform issues short-lived tokens automatically. Storing a key in an environment variable (A) still leaves a secret to manage. Embedding the key in source code (C) is a critical security violation. Key rotation (D) improves hygiene but doesn't remove the need to store the key somewhere.

3. 3. Contoso Health is deploying a patient triage chatbot. During testing, reviewers found that some model responses included offensive content. The company needs to prevent harmful outputs before they reach users while still allowing medically relevant discussions. What should the AI engineer implement?

   • A. Increase the model's temperature parameter to make responses more conservative
   • B. Implement Azure AI Content Safety content filters and blocklists within Microsoft Foundry(correct)
   • C. Add a post-processing step that replaces every response longer than 200 tokens
   • D. Switch from GPT-4 to a smaller model with fewer parameters

   Explanation: Azure AI Content Safety content filters, configurable in Microsoft Foundry, detect and block harmful content categories (hate, violence, sexual, self-harm) at configurable severity thresholds and allow custom blocklists for domain-specific terms. Lowering temperature (A) affects randomness, not safety. Truncating long responses (C) would cut off valid medical answers. Switching to a smaller model (D) does not add safety guarantees.

4. 4. An Azure AI engineer is setting up monitoring for a Microsoft Foundry Service resource. Which TWO actions should be taken to effectively monitor the service and manage costs? Choose 2.

   • A. Configure Azure Monitor diagnostic settings to send metrics and logs to a Log Analytics workspace(correct)
   • B. Set up Azure Cost Management budgets and alerts for the resource group(correct)
   • C. Enable debug logging inside each model's system prompt
   • D. Create a separate Azure subscription for every AI project to isolate billing
   • E. Delete the resource and recreate it monthly to reset usage counters

   Explanation: Azure Monitor diagnostic settings capture request counts, latency, errors, and token usage, enabling operational insight (A). Azure Cost Management budgets provide spending alerts before overage occurs (B). Embedding debug logs in a system prompt (C) adds tokens to every request but doesn't give infrastructure-level observability. A separate subscription per project (D) fragments governance. Deleting and recreating the resource (E) destroys history and configurations.

5. 5. An engineer needs to deploy an Azure AI service container to an edge location that has intermittent internet connectivity. The container must operate without a persistent connection to Azure. Which configuration requirement must be satisfied?

   • A. The container must be deployed with a GPU instance type in the edge environment
   • B. The container must be configured with a valid ApiKey and Billing endpoint so it can periodically report usage to Azure(correct)
   • C. The container requires a public IP address reachable from the Azure portal
   • D. The container must use a VNet-integrated Azure Kubernetes Service cluster only

   Explanation: Azure AI service containers in disconnected or edge scenarios still require the ApiKey and Billing environment variables so they can periodically phone home to Azure for metering (even if connectivity is intermittent). A GPU is not required for many AI service containers (A). A public IP is not required (C). Kubernetes is one deployment option but not the only one (D).

6. 6. A company's Responsible AI review found that a customer-facing sentiment analysis model consistently mis-classifies messages from non-native English speakers as negative. Which Responsible AI principle is most directly violated, and what is the primary remediation?

   • A. Reliability — retrain the model with more data from the same demographic group that is currently overrepresented
   • B. Fairness — audit training data for demographic imbalance and evaluate model performance across language subgroups, then mitigate disparities(correct)
   • C. Privacy — anonymize all input text before it reaches the model
   • D. Transparency — publish the model's F1 score in the company's annual report

   Explanation: Systematically worse performance for a specific demographic group (non-native English speakers) is a fairness violation. The remediation is to measure performance disaggregated by subgroup, identify the data gap, and apply bias mitigation techniques. Reliability (A) refers to consistent performance under valid conditions, not bias. Privacy (C) addresses data protection, not differential outcomes. Transparency (D) is about disclosure but does not fix the underlying disparity.

7. 7. A developer needs to ground a GPT-4 model's responses in a company's proprietary product catalog stored in Azure Blob Storage. The catalog is too large to include in a single context window. Which approach should the developer use?

   • A. Fine-tune the GPT-4 model on the product catalog text
   • B. Implement a Retrieval-Augmented Generation (RAG) pattern by indexing the catalog in Azure AI Search and retrieving relevant chunks at query time(correct)
   • C. Include the entire catalog as the system prompt on every API call
   • D. Use the DALL-E model to convert the catalog to images the model can read

   Explanation: RAG indexes the large catalog in Azure AI Search and retrieves only the most relevant document chunks to include in the prompt, staying within context limits while using current data. Fine-tuning (A) bakes knowledge into model weights at training time and is expensive to update. Including the entire catalog in the system prompt (C) would exceed the context window. DALL-E (D) generates images and cannot read documents.

8. 8. Adventure Works wants to track the quality of responses generated by its Azure OpenAI-powered support assistant over time. Product managers want to see whether the AI's answers degrade after model updates. Which capability in Microsoft Foundry should the engineer enable?

   • A. Enable prompt flow tracing and collect user feedback to monitor response quality trends(correct)
   • B. Increase max_tokens to generate longer responses after each model update
   • C. Add a fixed disclaimer to every AI response
   • D. Deploy a second identical model and compare their billing costs

   Explanation: Microsoft Foundry's prompt flow tracing captures execution details for every run, and collecting user feedback (thumbs up/down or explicit ratings) enables trend analysis of response quality over time. Increasing max_tokens (B) affects length, not quality measurement. Adding a disclaimer (C) doesn't measure quality. Comparing billing costs (D) measures spend, not answer quality.

9. 9. An engineer is building a generative AI application using Microsoft Foundry SDK. The team wants to version-control their prompt templates alongside application code. Which approach is recommended?

   • A. Store prompt templates as hardcoded strings in the application's main source file
   • B. Save prompt templates as separate files (e.g., .prompty or Jinja2 templates) checked into the Git repository alongside the application code(correct)
   • C. Regenerate the system prompt using DALL-E each time the application starts
   • D. Store prompts in an Azure Table Storage row with no version history

   Explanation: Microsoft Foundry SDK supports .prompty files and Jinja2 template formats that can be stored in Git, enabling the same branching, diffing, and code review workflows applied to application code. Hardcoded strings (A) are difficult to review and version. DALL-E (C) is an image-generation model and cannot create text prompts. Azure Table Storage without version history (D) loses the change audit trail.

10. 10. Fabrikam is running a high-traffic document summarization service on Azure OpenAI. During peak hours, users experience high latency. The engineering team has already set max_tokens conservatively. What should they configure to reliably handle sustained high-throughput demand?

    • A. Increase the temperature parameter to allow faster text sampling
    • B. Switch to the DALL-E deployment for text tasks to offload GPT-4 traffic
    • C. Provision a Provisioned Throughput Unit (PTU) deployment of the model to reserve dedicated capacity(correct)
    • D. Add a retry loop with exponential backoff and accept the latency as expected

    Explanation: Provisioned Throughput Units (PTUs) reserve dedicated model capacity in Azure OpenAI, guaranteeing consistent throughput and low latency under high-volume workloads instead of competing with other tenants on shared capacity. Temperature (A) controls randomness in output, not request throughput. DALL-E (B) generates images, not text. A retry loop with backoff (D) is a resilience pattern but doesn't increase actual capacity.

11. 11. An engineer wants to adapt a base GPT-4 model to respond in a highly specific medical terminology style that few-shot prompting cannot reliably achieve. The team has thousands of labeled example conversations in the correct style. What is the most appropriate approach?

    • A. Increase the system prompt length until the desired tone is established
    • B. Fine-tune the generative model using the labeled example conversations(correct)
    • C. Use Azure AI Search to index the example conversations and retrieve them at query time
    • D. Deploy a separate Azure AI Language model for intent recognition

    Explanation: Fine-tuning adjusts model weights using the labeled training data, enabling the model to reliably adopt the specific style, tone, and terminology without needing large few-shot examples in every prompt. Longer system prompts (A) still use few-shot inference and consume tokens every call. RAG via Azure AI Search (C) retrieves information but does not change the model's generation style. An intent recognition model (D) classifies intent but does not generate styled medical responses.

12. 12. A developer is building a customer support agent using the Microsoft Foundry Agent Service. The agent needs to look up order status from an internal REST API. How should the developer expose this capability to the agent?

    • A. Include all possible order records directly in the agent's system prompt
    • B. Define a tool (function) that wraps the REST API call and register it with the agent so the model can invoke it when needed(correct)
    • C. Create a separate GPT-4 deployment dedicated to order queries
    • D. Store order records in Azure Blob Storage and let the agent read blobs directly

    Explanation: The Microsoft Foundry Agent Service supports tool/function calling, where the developer defines a function schema that wraps the REST API; the model decides when to call it and passes the appropriate parameters. Embedding all order records in the system prompt (A) is not scalable. A dedicated GPT-4 deployment (C) still cannot call the internal API without tool definitions. Direct blob reads (D) require file-format parsing and bypass the structured API contract.

13. 13. A company wants to build a complex research assistant where one orchestrator agent breaks a research question into sub-tasks and delegates them to specialist agents (a web-search agent and a document-analysis agent). The agents must coordinate their outputs before returning a consolidated answer. Which implementation pattern should the engineer use?

    • A. Create a single GPT-4 model with a very long system prompt describing all three agent roles
    • B. Implement a multi-agent orchestration solution using Microsoft Agent Framework with an orchestrator agent routing tasks to specialist sub-agents(correct)
    • C. Use Azure Logic Apps to schedule each agent to run independently on a timer
    • D. Deploy three separate Azure AI Search indexes and merge results in the client application

    Explanation: Microsoft Agent Framework supports multi-agent orchestration where a controller agent decomposes tasks and routes them to specialist agents, then aggregates their outputs. A single monolithic prompt (A) cannot parallelize sub-tasks or maintain clean separation of concerns. Logic Apps with timers (C) are event-driven workflows without dynamic orchestration logic. Merging search indexes (D) retrieves documents but does not reason or orchestrate agent behaviour.

14. 14. A warehouse management system needs to automatically count the number of boxes and identify their positions in images captured by ceiling cameras. Which Azure AI Vision capability should the developer use?

    • A. Image classification
    • B. Object detection(correct)
    • C. Optical character recognition (OCR)
    • D. Facial detection

    Explanation: Object detection identifies the presence and bounding-box positions of multiple objects (boxes) in an image, returning both class labels and coordinates. Image classification (A) assigns a single label to the whole image without locating individual objects. OCR (C) extracts text characters. Facial detection (D) is specific to human faces.

15. 15. Contoso Retail trains a custom vision model to classify product images as 'Damaged' or 'Undamaged' on a manufacturing line. After publishing the model, precision is 95% but recall is 68%. The quality team considers missed damaged products (false negatives) more costly than false alarms. What should the engineer adjust?

    • A. Lower the probability threshold for classifying an image as 'Damaged' to increase recall at the cost of some precision(correct)
    • B. Increase the probability threshold to improve precision further
    • C. Switch from classification to an object detection model
    • D. Retrain the model with fewer training images to reduce overfitting

    Explanation: When false negatives (missing damaged products) are more costly, recall must be increased. Lowering the classification threshold causes the model to flag more images as 'Damaged', reducing false negatives at the expense of some additional false alarms (lower precision). Raising the threshold (B) would improve precision but further reduce recall. Switching to object detection (C) changes the task type without addressing the threshold problem. Fewer training images (D) would likely worsen both metrics.

16. 16. An engineer needs to extract printed and handwritten text from scanned medical forms and convert it into a searchable digital format. Which TWO Azure AI Vision capabilities in Foundry Tools should be used? Choose 2.

    • A. Extract text from images using Azure Vision in Foundry Tools(correct)
    • B. Convert handwritten text using Azure Vision in Foundry Tools(correct)
    • C. Use Azure AI Video Indexer to extract text frames from video
    • D. Use the DALL-E model to recreate clean versions of the forms
    • E. Train a custom object detection model on the forms

    Explanation: Azure Vision in Foundry Tools provides both printed text OCR (A) and handwriting recognition (B) in a unified Read API call, making both capabilities the correct choice for mixed printed/handwritten content. Azure AI Video Indexer (C) processes video streams, not static scanned images. DALL-E (D) generates new images and cannot transcribe text. Custom object detection (E) locates objects within images but does not transcribe text characters.

17. 17. A developer needs to automatically identify names of people, organisations, and locations mentioned in customer feedback text using Azure AI Language. Which feature should be used?

    • A. Key phrase extraction
    • B. Named entity recognition (NER)(correct)
    • C. Sentiment analysis
    • D. Language detection

    Explanation: Named entity recognition (NER) identifies and classifies entities such as people, organisations, and locations within text. Key phrase extraction (A) returns salient terms but does not classify them by entity type. Sentiment analysis (C) scores emotional tone. Language detection (D) identifies the language of the text.

18. 18. A healthcare portal processes patient messages that may contain phone numbers, email addresses, and national ID numbers. Before storing messages in a shared analytics database, the team must redact this information. Which Azure AI Language feature should they use?

    • A. Sentiment analysis to detect negative messages and flag them for manual review
    • B. Personally identifiable information (PII) detection to identify and redact sensitive entities from the text(correct)
    • C. Entity linking to map entities to a knowledge base
    • D. Custom text classification to label each message category

    Explanation: The Azure AI Language PII detection feature identifies categories such as phone numbers, email addresses, and government IDs, returning their positions and categories so the application can redact or replace them before storage. Sentiment analysis (A) measures tone and cannot locate or mask PII. Entity linking (C) connects mentions to external knowledge bases, not redaction. Custom text classification (D) assigns categorical labels, not PII masking.

19. 19. A call-centre application needs to transcribe spoken customer calls in real time and route them based on the caller's intent (e.g., 'cancel subscription', 'billing question'). Which TWO Azure AI Speech in Foundry Tools capabilities must the engineer combine?

    • A. Speech-to-text for real-time transcription, combined with intent recognition to classify the caller's goal(correct)
    • B. Text-to-speech to read responses back to the caller
    • C. Custom speech model training to improve transcription accuracy for call-centre vocabulary
    • D. DALL-E image generation to visualise the caller's sentiment

    Explanation: The solution requires speech-to-text to convert audio to text in real time, then intent recognition (via Azure AI Speech or Azure AI Language) to classify what the caller wants. Text-to-speech (B) outputs spoken audio and is a different direction. Custom speech training (C) improves accuracy but does not add routing logic. DALL-E (D) generates images and is entirely unrelated to call routing.

20. 20. An international e-commerce company needs to translate customer reviews submitted in any of 50 languages into English for analysis. Some reviews contain product-specific brand names that must not be translated. Which feature of Azure Translator in Foundry Tools should the engineer configure?

    • A. Use the 'to' parameter set to 'en' and rely on default translation behaviour
    • B. Configure a custom glossary (dynamic dictionary) that maps brand names to themselves so they are preserved verbatim(correct)
    • C. Enable profanity filtering to protect brand integrity
    • D. Use a custom neural machine translation model trained exclusively on English text

    Explanation: Azure Translator supports dynamic dictionary entries (inline glossaries) that instruct the engine to pass specific terms through untranslated. This is exactly the right mechanism for preserving brand names across languages. Default translation (A) will attempt to translate or transliterate brand names. Profanity filtering (C) is unrelated to brand name preservation. A model trained only on English (D) cannot translate from 50 source languages.

21. 21. A company is building a custom question-answering bot using Azure AI Language. After initial deployment, users report that the bot does not handle variations of the same question well and gives robotic responses to greetings. Which TWO improvements should the engineer make? Choose 2.

    • A. Add alternate phrasing to existing question-answer pairs to improve matching on question variants(correct)
    • B. Add chit-chat to the knowledge base to handle conversational greetings and small talk(correct)
    • C. Retrain the knowledge base with the system prompt from the GPT-4 model
    • D. Export and delete the knowledge base, then rebuild it from scratch
    • E. Increase the confidence threshold to 99% to reduce incorrect answers

    Explanation: Adding alternate phrasing to QnA pairs (A) teaches the model to match semantically equivalent wordings of the same question. Adding chit-chat (B) is a built-in Azure AI Language feature that inserts pre-built conversational responses for greetings and social exchanges. GPT-4 system prompts (C) are not a valid input for the QnA knowledge base. Deleting and rebuilding (D) would erase all existing content. A 99% confidence threshold (E) would cause the bot to refuse most questions rather than improving quality.

22. 22. A legal firm wants to make thousands of scanned contracts searchable by clause type and legal party names. The documents are stored in Azure Blob Storage. What is the correct sequence of components to configure in Azure AI Search?

    • A. Create an index → run the indexer → define the skillset → add a data source
    • B. Provision an Azure AI Search resource → create a data source pointing to Blob Storage → define an index → attach a skillset → create and run an indexer(correct)
    • C. Provision an Azure AI Search resource → deploy a GPT-4 model → use the model to read all blobs
    • D. Create a data source → publish the knowledge base → run semantic ranking

    Explanation: The standard Azure AI Search pipeline is: provision the resource, create a data source connector to Blob Storage, define the index schema, attach an enrichment skillset (for entity extraction, key phrase extraction, etc.), then create and run an indexer to orchestrate the pipeline. The sequence in (A) is out of order. GPT-4 deployment (C) is not part of the AI Search indexing pipeline. A 'knowledge base' (D) refers to the QnA service, not AI Search.

23. 23. Northwind Traders processes thousands of supplier invoices monthly. Each invoice has a different layout. The company needs to extract vendor name, invoice number, line items, and total amount automatically. Which Azure service and model type is most appropriate?

    • A. Azure AI Vision image tagging to identify each invoice as 'invoice type'
    • B. Azure Document Intelligence in Foundry Tools with the prebuilt invoice model to extract standard invoice fields(correct)
    • C. Azure AI Language key phrase extraction applied to the raw invoice text
    • D. Azure AI Search full-text indexing of the scanned invoice PDFs

    Explanation: Azure Document Intelligence includes a prebuilt invoice model that understands common invoice structures and extracts semantic fields (vendor name, invoice number, line items, totals) without layout-specific configuration. Vision image tagging (A) classifies the document type but does not extract structured fields. Key phrase extraction (C) returns salient phrases but lacks invoice-specific field mapping. Full-text search (D) makes documents searchable but does not extract structured data fields.

24. 24. An engineer needs to add a custom enrichment step to an Azure AI Search skillset that calls an internal REST API to classify documents by business department. The built-in skills do not cover this classification logic. Which TWO steps are required to integrate this custom logic? Choose 2.

    • A. Create a custom skill definition that specifies the URI of the REST endpoint and the expected input/output schema(correct)
    • B. Include the custom skill in the skillset definition so the indexer calls it during enrichment(correct)
    • C. Replace the entire Azure AI Search service with a custom Elasticsearch cluster
    • D. Add the classification field directly to the source documents in Blob Storage before indexing
    • E. Retrain the Azure AI Search neural ranking model with the department taxonomy

    Explanation: Custom skills in Azure AI Search are REST endpoints; the engineer defines the skill's URI, inputs, and outputs in a JSON definition (A), then references the custom skill within the skillset so the indexer invokes it during document enrichment (B). Replacing AI Search with Elasticsearch (C) is an unnecessary infrastructure change. Pre-labelling blobs (D) bypasses the enrichment pipeline purpose. Retraining the neural ranking model (E) affects result ordering, not field extraction.

25. 25. A media company processes thousands of hours of archived video content. They need to automatically extract spoken dialogue, identify speakers, detect on-screen topics, and surface key moments for editorial review. Which Azure service is designed for this multi-modal video enrichment scenario?

    • A. Azure AI Document Intelligence, using the prebuilt layout model on video frame screenshots
    • B. Azure AI Video Indexer to extract insights such as transcripts, speaker identification, topics, and keyframes from video(correct)
    • C. Azure AI Speech speech-to-text only, with no video-level insights
    • D. Azure Cognitive Search semantic ranking applied to an index of video metadata

    Explanation: Azure AI Video Indexer is purpose-built for video AI enrichment, providing automatic transcription, speaker diarization, topic detection, keyframe extraction, named entity recognition within content, and more — all from video files. Document Intelligence (A) processes documents, not video streams. Speech-to-text alone (C) transcribes audio but provides no visual or topic-level insights. Semantic ranking in AI Search (D) improves search relevance for already-indexed text, but does not process video.