# Free AI-500 practice questions

> 10 free AI-500 practice questions with the correct answer and an explanation.

Page: https://devopscodelab.com/exams/microsoft-ai-500/free-practice-questions/
Practice exams: https://devopscodelab.com/exams/microsoft-ai-500/

### Question 1 of 10

A bank is designing a multi-agent system in which a triage agent routes requests to a loans agent and a fraud agent built by another team on a different framework and hosted in another Azure tenant. The teams want a standard for agent discovery, task delegation, and streaming status updates between the agents. Which protocol should you specify between the agents?

- A. Model Context Protocol (MCP), exposing each remote agent as a tool with a JSON schema
- B. Azure Service Bus topics carrying custom JSON messages that each team parses in its own format
- C. Agent2Agent (A2A), using agent cards for discovery and task-based messages between independent agents
- D. OpenAPI 3.0 specifications for each agent's REST endpoints, called from the triage agent's function tools

**Answer:** C

**Explanation:** A2A is designed for communication between independent, opaque agents across frameworks and organizations, with agent cards for discovery and task lifecycle messages including streaming updates. MCP standardizes how an agent connects to tools and context, not peer agent delegation. OpenAPI endpoints and custom Service Bus messages can move data but give no shared model for discovery, task state, or agent capabilities, so each team would reinvent the contract.

### Question 2 of 10

You are defining a new procurement agent that can read supplier catalogs and draft purchase orders. Company policy says no agent can commit spend on its own. Which specification BEST enforces this boundary?

- A. Add a system prompt instruction that tells the agent never to submit purchase orders and to explain the approval policy if a user asks
- B. Fine-tune the model on examples in which the agent refuses to submit orders, then grant it the ERP submit permission as a fallback
- C. Give the agent the full ERP API with a Contributor role, and audit its orders weekly
- D. Give it read-only catalog and draft-only order tools, and route submission through a human approval step

**Answer:** D

**Explanation:** Autonomy boundaries are enforced most reliably through the tools and permissions an agent is given, so read-only and draft-only tools plus a human-completed submission step make committing spend impossible for the agent. Prompt instructions and fine-tuned refusals are behavioral guidance that can be bypassed by prompt injection or model error while the permission still exists. A weekly audit detects problems only after money has been committed.

### Question 3 of 10

Retrieval quality in a multi-agent knowledge base is poor for questions written in German, although the indexed documents are in both German and English. The index uses an English-only embedding model chosen two years ago. What should you change, and what else must happen?

- A. Translate every German query into English with a translation agent before retrieval, and keep the English-only embeddings as they are
- B. Increase the vector dimensions of the existing index from 1,536 to 3,072 so that the English-only model captures more language detail
- C. Use a multilingual embedding model for queries only and keep the existing document vectors
- D. Switch to a multilingual embedding model and re-embed the whole corpus, since old and new vectors are not comparable

**Answer:** D

**Explanation:** A multilingual embedding model improves cross-language retrieval, but vectors from different models live in different spaces, so the whole corpus must be re-embedded. Embedding only queries with the new model makes them incomparable with stored vectors. Query translation helps partly but still misses German-only nuance in documents embedded by an English-only model, and changing index dimensions does not change what the old model can represent.

### Question 4 of 10

A legal research agent built on gpt-4.1 follows the firm's citation format correctly only about 70% of the time, even with detailed instructions and examples. You have 3,000 high-quality, reviewed input/output pairs, and the format changes about once a year. What should you do?

- A. Switch to a larger reasoning model and remove the citation examples from the prompt to give the model more freedom
- B. Supervised fine-tuning on the reviewed pairs, held-out evaluation, and a refresh when the format changes
- C. Add all 3,000 reviewed pairs to the system prompt as few-shot examples on every request to the legal research agent
- D. Fine-tune the base model again every week with the newest 50 production conversations so that it adapts to recent requests

**Answer:** B

**Explanation:** Consistent formatting that prompting cannot achieve is a good fit for supervised fine-tuning on curated data, with held-out evaluation and a retraining cadence tied to how often the format changes. Weekly retraining on unreviewed production logs risks learning mistakes and wastes cost. Thousands of few-shot examples would exceed practical prompt sizes, and a bigger model without examples does not target the formatting gap.

### Question 5 of 10

You are migrating agents from Semantic Kernel and AutoGen to a single framework recommended by Microsoft for new multi-agent development in Python and .NET. Which framework should you target?

- A. Microsoft Bot Framework SDK v4
- B. Azure Logic Apps Standard
- C. Prompt flow in Foundry
- D. Microsoft Agent Framework

**Answer:** D

**Explanation:** Microsoft Agent Framework is the successor that combines concepts from Semantic Kernel and AutoGen for building agents and multi-agent workflows in Python and .NET. The Bot Framework SDK targets conversational bots and is not the recommended agent framework. Prompt flow is a development tool for LLM flows, and Logic Apps is a workflow integration service rather than an agent SDK.

### Question 6 of 10

A multi-agent system stores long-term memories that include customers' personal data. Compliance requires encryption at rest with company-managed keys, a 13-month retention limit, and the ability to delete one customer's memories on request. Which memory strategy meets these requirements?

- A. Customer-managed-key encryption, records keyed by customer, TTL expiry, and per-customer deletion
- B. Memories kept only in each agent's in-process cache, lost on restart
- C. A shared memory document per agent containing all customers' facts, protected with Microsoft-managed encryption keys
- D. Memories embedded inside the model through monthly fine-tuning, with a new model version deployed each month

**Answer:** A

**Explanation:** Customer-managed key encryption, per-customer keys for records, TTL expiry, and a targeted delete operation satisfy encryption, retention, and erasure requirements. Fine-tuning bakes personal data into weights where targeted deletion is impractical. In-process caches lose memories unpredictably, and a shared document per agent cannot delete one customer's data cleanly and uses Microsoft-managed rather than company-managed keys.

### Question 7 of 10

A planning agent in Microsoft Foundry often chooses the wrong tool or passes incorrect parameters, even though its final answers sound fluent. Which built-in evaluator should you add to measure this problem directly?

- A. Groundedness evaluator
- B. Fluency evaluator
- C. Tool call accuracy evaluator
- D. Response completeness evaluator

**Answer:** C

**Explanation:** The tool call accuracy evaluator assesses whether the agent selected appropriate tools and supplied correct parameters, which is exactly the observed failure. Fluency measures how readable the text is, which already looks fine here. Response completeness checks whether the answer covers the expected content, and groundedness checks whether an answer is supported by provided context; neither inspects which tool was called or with what arguments.

### Question 8 of 10

You want Foundry tracing to show, for each agent run, the prompts, completions, token counts, and tool calls. Which configuration step is required?

- A. Turn on Azure Monitor VM insights for the hosts that run the agent containers
- B. Enable Microsoft Defender for Cloud on the subscription that contains the Foundry resource
- C. Connect Application Insights to the Foundry project and instrument the agent with OpenTelemetry-based tracing
- D. Enable diagnostic settings on the Azure OpenAI resource and send its request and audit logs to a storage account and Log Analytics

**Answer:** C

**Explanation:** Foundry tracing stores traces in an Application Insights resource connected to the project, and the agent code is instrumented with OpenTelemetry-based tracing to capture prompts, completions, tokens, and tool calls. Resource diagnostic logs record control and request metadata, not agent spans. VM insights monitor host performance, and Defender for Cloud provides security posture rather than execution traces.

### Question 9 of 10

Your security team wants to find prompt injection and jailbreak weaknesses in an agent before it reaches production, as part of the development cycle. Which Foundry capability should you use?

- A. Microsoft Defender for Cloud regulatory compliance dashboard, which scores the subscription against Azure security benchmarks
- B. The AI Red Teaming Agent, which runs automated adversarial probes and reports attack success rate per risk category
- C. The groundedness evaluator, run on a sample of production conversations after the agent has launched
- D. Application Insights availability tests that call the agent endpoint every five minutes from several regions

**Answer:** B

**Explanation:** The AI Red Teaming Agent automates adversarial probing, including attack strategies that try to bypass safety behavior, and reports attack success rate so issues are found before release, which is shift-left security. A compliance dashboard checks resource configuration, not model behavior. Availability tests check uptime, and groundedness evaluation in production is neither adversarial nor shifted left.

### Question 10 of 10

A document-processing agent reads emails and web pages that might contain hidden instructions targeting the agent. Which guardrail should you apply to this retrieved content?

- A. A blocklist of profanity terms applied to the user's typed messages in the chat interface
- B. Protected material detection for text on the agent's final output to catch copyrighted song lyrics and articles
- C. Groundedness detection on the model's answer to verify that claims are supported by the source documents
- D. Prompt Shields for document (indirect) attacks on the untrusted content before it is passed to the model

**Answer:** D

**Explanation:** Indirect prompt injection hides instructions in content the agent processes, and Prompt Shields document attack detection is designed to flag it before the model acts. Protected material detection looks for copyrighted text in outputs. A profanity blocklist on typed messages does not inspect retrieved documents, and groundedness detection checks factual support, not embedded instructions.

Ready for a timed run? Take the free 15-question AI-500 mini exam: https://devopscodelab.com/app/exams/microsoft-ai-500/mini/

DevOpsCode Lab is not affiliated with Microsoft. AI-500 and related names are trademarks of their owners. We write every question ourselves; these are not real exam questions.
