Azure AI-200 Sample Questions with Explanations

These original sample questions are designed to help you check how the exam topics appear in decision-style prompts. They are not taken from the live exam.

Use these sample questions as a guided self-assessment for Microsoft AI Cloud Developer route (AI-200) topics such as AI application architecture, model-flow integration, retrieval, tool calls, deployment choices, monitoring, identity, and responsible AI controls. The prompts emphasize developer decisions in cloud AI systems.

Where these questions fit in the AI-200 guide

The sample set below is part of the Microsoft AI-200 guide path:

• Microsoft certification hub
• Microsoft Azure guides
• AI-200 exam guide
• AI-200 cheat sheet
• AI-200 study plan
• AI-200 FAQ
• AI-200 resources
• AI-200 glossary

AI-200 cloud developer sample questions

Work through each prompt before opening the explanation. Strong answers usually choose the architecture that makes model behavior observable, secure, and maintainable.

Question 1

Topic: Moving a prototype to production

A developer built an AI prototype in a notebook. The app now needs production deployment with repeatable prompts, versioned configuration, monitored latency, and a rollback path if a prompt change causes poor responses. What should the team prioritize?

• A. Keep all prompt changes inside the notebook because it is faster to edit.
• B. Increase the model temperature so production responses are less repetitive.
• C. Move the model flow into a deployable application pipeline with versioned prompt/config artifacts, telemetry, staged rollout, and rollback controls.
• D. Disable logging because prompts may contain sensitive data.

Best answer: C

Explanation: The scenario is about operationalizing an AI application, not just improving response style. Versioning, telemetry, staged rollout, and rollback make prompt and configuration changes manageable in production.

Why the other choices are weaker:

• A keeps the system in a fragile prototype workflow.
• B changes model behavior but does not solve deployment control.
• D removes observability; sensitive data should be handled with redaction, access control, and retention rules instead.

What this tests: Moving from experimentation to controlled AI application deployment.

Related topics: Deployment; Prompt versioning; Monitoring; Rollback

Question 2

Topic: Tool call guardrails

An AI assistant can look up customer orders and submit refund requests. The business wants lookup to be automatic, but refund submission must be limited to eligible orders and logged for audit. Which design is best?

• A. Give the model direct access to the refund API with a shared administrator credential.
• B. Put refund rules in the user prompt and trust the model to follow them.
• C. Expose lookup and refund as separate tools, validate refund inputs against eligibility rules, use scoped identity, and log every tool call.
• D. Remove lookup access so the assistant can only give generic refund policy answers.

Best answer: C

Explanation: A production AI app should separate tools by risk and put deterministic checks around sensitive actions. Scoped identity, validation, and logging make the action path safer and reviewable.

Why the other choices are weaker:

• A violates least privilege and weakens accountability.
• B treats a policy requirement as prompt text rather than an enforceable control.
• D avoids risk but fails the lookup requirement.

What this tests: Designing controlled tool use for cloud AI applications.

Related topics: Tool use; Validation; Least privilege; Audit logging

Question 3

Topic: Retrieval freshness

A policy assistant sometimes returns guidance from old procedure documents. The team confirms the answer generator is working but the retrieved passages include retired documents. What is the best fix?

• A. Increase the maximum response length so the model can include more caveats.
• B. Ask users to include the current date in every question.
• C. Fine-tune the model on the old and new policies together so it learns both versions.
• D. Add metadata such as approval status and effective date to the index, filter retired content at retrieval time, and re-index when policy status changes.

Best answer: D

Explanation: The failure is in retrieval governance, not output length. Metadata filters and index updates keep retired content out of the context passed to the model.

Why the other choices are weaker:

• A may make wrong answers longer.
• B pushes a system control onto users and still may retrieve retired documents.
• C makes version control less explicit and does not guarantee current policy selection.

What this tests: Troubleshooting retrieval quality by controlling source metadata and index lifecycle.

Related topics: Retrieval; Indexing; Freshness; Governance

Tech Exam Lexicon and IT Mastery are independent study tools. They are not affiliated with, endorsed by, or sponsored by the exam vendor.