AI-102: Azure AI Engineer Associate Practice Exam

What the AI-102 exam covers

• Plan and Manage an Azure AI Solution143 questions
• Implement Content Moderation Solutions120 questions
• Implement Computer Vision Solutions129 questions
• Implement NLP Solutions146 questions
• Implement Knowledge Mining and Document Intelligence137 questions
• Implement Generative AI Solutions136 questions
• Implement Agentic Solutions35 questions

Free AI-102 sample questions

A sample of 10 questions with answers and explanations. Sign up free to practice all 846.

1. Question 1Plan and Manage an Azure AI Solution

   You are deploying an Azure AI Services multi-service resource. You need to restrict network access so that only your virtual network can reach the endpoint. What should you configure?

   • AConfigure a private endpoint and disable public network accessCorrect
   • BEnable customer-managed keys for encryption
   • CEnable a managed identity on the resource
   • DRotate the subscription keys
   ✓ Correct answer: A

   Private endpoints create a network interface within your virtual network that connects securely to Azure AI Services, ensuring traffic never traverses the public internet. Disabling public network access prevents any connections from outside your VNet, providing complete network isolation. This is the standard Azure approach for restricting access to managed services to only your own network infrastructure.

   Why the other options are wrong

   • BEnable customer-managed keys for encryption is incorrect because encryption configuration does not restrict network access; it only secures data at rest and is separate from network isolation requirements.
   • CEnable a managed identity on the resource is incorrect because managed identities are for authentication and authorization, not for controlling which networks can reach the endpoint.
   • DRotate the subscription keys is incorrect because key rotation manages authentication credentials but does not restrict network access to the endpoint.
2. Question 2Implement Content Moderation Solutions

   A developer is configuring Azure AI Content Safety thresholds for Fabrikam Inc's children's educational platform. The platform requires the strictest possible content filtering. How should the developer configure the severity thresholds?

   • ASet the threshold to 0 so all content flagged at any severity level is blockedCorrect
   • BDisable content moderation and rely on user reporting
   • CSet the threshold to 6 so only the most severe content is blocked
   • DSet the threshold to 4 for a moderate filtering level
   ✓ Correct answer: A

   For a children's educational platform requiring the strictest content filtering, setting the threshold to 0 ensures maximum protection.

   Why the other options are wrong

   • BDisable content moderation and rely on user reporting is incorrect because it leaves children exposed to harmful content without automated protection, which is inappropriate for a platform designed for children.
   • CThe Azure AI Content Safety severity scale ranges from 0 (safe) to 6 (most severe), and lower threshold values mean more content gets blocked. A threshold of 0 captures all flagged content regardless of severity level, providing the most restrictive filtering appropriate for protecting minors from any potentially harmful material.
   • DSet the threshold to 6 is incorrect because this only blocks the most extreme content while allowing mild to moderate harmful content through, contradicting the requirement for strictest filtering. Set the threshold to 4 for a moderate filtering level is incorrect because it represents only moderate filtering, not the strictest level needed for a children's platform.
3. Question 3Implement Computer Vision Solutions

   You are building a document processing solution that must extract data from invoices with varying layouts from hundreds of different vendors. Which Azure service is most appropriate?

   • AAzure Form Recognizer custom model trained on every vendor's invoice
   • BAzure AI Document Intelligence prebuilt invoice modelCorrect
   • CAzure Cognitive Search with OCR skill
   • DAzure AI Vision OCR
   ✓ Correct answer: B

   Azure AI Document Intelligence provides prebuilt models specifically designed for common document types, including invoices. The prebuilt invoice model is trained on thousands of invoice examples from various vendors and can extract key information such as invoice numbers, dates, totals, and line items regardless of the specific vendor's layout. This approach is more efficient than training custom models for each vendor.

   Why the other options are wrong

   • AAzure Form Recognizer custom model trained on every vendor's invoice is incorrect because training a custom model for each vendor would be impractical with hundreds of different vendors and would require continuous retraining as new vendor formats appear.
   • CAzure Cognitive Search with OCR skill is incorrect because Cognitive Search is designed for indexing and searching document content, not for extracting structured invoice data with field-level accuracy.
   • DAzure AI Vision OCR is incorrect because basic OCR extracts all text without understanding invoice structure or extracting key fields like amounts and dates.
4. Question 4Implement NLP SolutionsSelect all that apply

   You are implementing Azure AI Translator for a global e-commerce application. Which two features can you use to ensure translation quality for domain-specific terminology? (Choose two.)

   • ADynamic dictionary to force specific term translations at request timeCorrect
   • BAzure AI Translator auto-detect to identify the source language
   • CAzure AI Language text analytics to validate translations
   • DCustom Translator with a trained model using your product catalogs as training dataCorrect
   ✓ Correct answer: A, D

   Both selected features directly target terminology quality. The dynamic dictionary forces exact translations of marked terms within individual requests, useful for a small set of must-be-correct phrases. Custom Translator goes further by training a model on your parallel catalog data so the entire vocabulary and style of your domain are learned and applied automatically. Together they cover both ad hoc and systematic terminology control. The other options handle language detection or analysis, not terminology accuracy.

   Why the other options are wrong

   • BAzure AI Translator auto-detect only identifies the source language of input text and does nothing to control how domain terms are translated.
   • CAzure AI Language text analytics analyzes text for sentiment, entities, or key phrases and is not a translation-quality mechanism.
5. Question 5Implement Generative AI Solutions

   You need to customize a GPT model's behavior for a specific task using your own training examples, but prompt engineering alone is insufficient. What should you do?

   • AFine-tune the model with a training dataset of example prompt-completion pairsCorrect
   • BUse a larger max_tokens value
   • CIncrease the temperature parameter
   • DAdd more few-shot examples to the system message
   ✓ Correct answer: A

   When prompt engineering reaches its limits for task-specific customization, fine-tuning adjusts model weights using domain-specific examples. This enables the model to learn task-specific patterns, terminology, and response styles that few-shot examples cannot achieve. Fine-tuning creates a specialized model that performs the target task consistently and accurately.

   Why the other options are wrong

   • BUse a larger max_tokens value is incorrect because larger token limits increase output length, not task-specific behavior.
   • CIncrease the temperature parameter is incorrect because higher temperature increases randomness, which worsens task-specific consistency.
   • DAdd more few-shot examples to the system message is incorrect because few-shot learning is part of prompt engineering, which the scenario states is already insufficient.
6. Question 6Implement Content Moderation Solutions

   A team is planning Moderation procedures for Implement Content Moderation Solutions. What should they prioritize?

   • AImplement role-based access control with least privilegeCorrect
   • BGrant full administrator access to all team members
   • CUse a single shared service account for the entire team
   • DDisable access controls for faster day-to-day workflows
   ✓ Correct answer: A

   Role-based access control (RBAC) with least privilege ensures that team members have only the minimum permissions necessary to perform their assigned functions. This principle reduces the attack surface by limiting the impact of compromised credentials or insider threats. Implementing distinct roles for different team responsibilities enables fine-grained access policies, simplifies audit trail analysis, and supports compliance requirements that mandate principle of least privilege. RBAC also facilitates delegation of specific tasks without granting unnecessary administrative rights.

   Why the other options are wrong

   • BGrant full administrator access to all team members is incorrect because universal administrator access violates the principle of least privilege, increases the risk impact of credential compromise, and makes it difficult to audit who made critical changes.
   • CUse a single shared service account for the entire team is incorrect because shared accounts eliminate individual accountability, prevent proper audit trail attribution, and increase the risk of unauthorized actions going undetected.
   • DDisable access controls for faster day-to-day workflows is incorrect because removing access controls exposes the system to unauthorized modifications, eliminates security boundaries, and creates compliance violations.
7. Question 7Plan and Manage an Azure AI Solution

   You are implementing an Azure AI solution and must ensure that the AI model treats all demographic groups equitably. Which responsible AI principle are you addressing?

   • AReliability and safety
   • BFairnessCorrect
   • CAccountability
   • DTransparency
   ✓ Correct answer: B

   Fairness is the responsible AI principle that explicitly addresses equitable treatment across demographic groups and population segments. When ensuring that an AI model treats all demographic groups equitably, you are implementing fairness by detecting and mitigating algorithmic bias, ensuring equal prediction accuracy across groups, and preventing discriminatory outcomes. This principle ensures that AI systems do not disadvantage or harm specific populations based on protected characteristics like race, gender, age, or socioeconomic status.

   Why the other options are wrong

   • AReliability and safety is incorrect because it addresses system stability and failure modes rather than demographic equity.
   • CAccountability is incorrect because it concerns transparency in decision-making and responsibility assignment rather than equitable treatment.
   • DTransparency is incorrect because it focuses on explaining decisions rather than ensuring those decisions are equitable across groups.
8. Question 8Implement Computer Vision Solutions

   You need to count the number of people entering and exiting a retail store entrance using a fixed overhead camera. Which Azure AI capability should you use?

   • AAzure AI Vision Image Analysis People Detection
   • BAzure AI Vision Spatial AnalysisCorrect
   • CAzure Custom Vision Object Detection
   • DAzure AI Video Indexer
   ✓ Correct answer: B

   Azure AI Vision Spatial Analysis is the specialized service designed for counting people and tracking movement patterns using fixed camera feeds. It provides specific cognitive operations for scenarios like entry/exit counting, zone occupancy analysis, and movement tracking. Spatial Analysis is purpose-built for fixed camera surveillance and occupancy analytics, unlike general vision APIs.

   Why the other options are wrong

   • AAzure AI Vision Image Analysis People Detection is incorrect because Image Analysis is designed for single-image analysis with general people detection; it does not provide the continuous frame analysis and sophisticated line/zone crossing tracking needed for persistent counting.
   • CAzure Custom Vision Object Detection is incorrect because while Custom Vision could detect people, it is designed for specific object recognition training and lacks the built-in movement tracking and occupancy analytics that Spatial Analysis provides.
   • DAzure AI Video Indexer is incorrect because Video Indexer is designed for video content analysis and insights like transcription and topics; it does not provide real-time spatial analytics for entry/exit counting.
9. Question 9Implement NLP Solutions

   You need to create a CLU entity that matches a specific pattern, such as a flight number format like 'AA1234' or 'UA567'. Which entity type should you use?

   • ALearned entity
   • BList entity
   • CPrebuilt entity
   • DRegex entityCorrect
   ✓ Correct answer: D

   Regex entities are designed to match patterns expressed as regular expressions. Flight number formats like 'AA1234' or 'UA567' follow specific structural patterns (two letters followed by numbers). A regex entity with pattern like ^[A-Z]{2}\d{3,4}$ matches this exact format reliably. This approach is more efficient and accurate than learned entities for pattern-based matching and does not require labeled training examples.

   Why the other options are wrong

   • ALearned entity is incorrect because learned entities require many labeled examples to recognize patterns and are less precise than regex matching for format-based entities.
   • BList entity is incorrect because list entities are for discrete value sets, not pattern-based formats that can have many valid variations.
   • CPrebuilt entity is incorrect because prebuilt entities recognize general categories like numbers or dates but cannot enforce specific flight number format constraints.
10. Question 10Implement Generative AI Solutions

    You are deploying an Azure OpenAI model and need to manage token consumption. Your application makes requests that sometimes produce very long outputs. Which parameter should you set to control the maximum length of generated responses?

    • Atemperature
    • Bmax_tokensCorrect
    • Ctop_p
    • Dpresence_penalty
    ✓ Correct answer: B

    The max_tokens parameter sets the upper bound on the number of tokens the model can generate in a single response. This parameter directly controls response length and prevents unbounded token consumption, making it essential for managing costs and controlling output scope. By configuring this parameter, you ensure that even if a prompt is designed to generate lengthy outputs, the model will terminate generation at the specified token limit.

    Why the other options are wrong

    • ATemperature is incorrect because temperature controls randomness in token selection, not output length.
    • CTop_p is incorrect because top_p (nucleus sampling) controls diversity through probability thresholding, not the maximum length.
    • DPresence_penalty is incorrect because presence_penalty adjusts the likelihood of introducing new topics, not the total token output limit.

AI-102 practice exam FAQ