Free AI-200 sample questions with answers
Here are five free AI-200 sample questions, one or more from each skill area, with the answer and the reasoning behind it. Try each one before opening the answer. These are original scenario questions written against the official skills outline, not questions from the real exam.
Question 1. A Python API runs in Azure Container Apps. It must scale out when messages pile up in an Azure Service Bus queue and scale to zero replicas when the queue is empty. What should you configure?
- A. An HTTP scale rule with a concurrency threshold
- B. A CPU utilisation scale rule
- C. A KEDA Azure Service Bus scale rule with minimum replicas set to 0
- D. A fixed replica count equal to the peak message rate
Show answer
Answer: C
Container Apps uses KEDA for event-driven scaling, and a Service Bus scale rule watches queue length and can scale to zero when minimum replicas is set to 0. An HTTP rule reacts to incoming requests, not queued messages. A CPU rule cannot scale to zero, because an idle app uses no CPU to trigger on. A fixed replica count removes scaling altogether.
Want more questions like this? Full AI-200 practice tests →
Question 2. You store product descriptions and their embeddings in Azure Cosmos DB for NoSQL. Users type a question, and the app must return the five most semantically similar products. Queries are slow and expensive. What should you do first?
- A. Switch the account to strong consistency
- B. Add a vector embedding policy and a vector index for the embedding path, and query with VectorDistance
- C. Replace the query with a LIKE filter on the description
- D. Increase the provisioned RU/s on the container
Show answer
Answer: B
Defining a vector embedding policy and a vector index on the embedding path lets Cosmos DB run VectorDistance queries efficiently instead of scanning every item. Strong consistency increases RU cost and does not help similarity search. A full-text LIKE query matches words, not meaning. Raising provisioned throughput pays for the inefficiency rather than removing it.
Want more questions like this? Full AI-200 practice tests →
Question 3. Orders are processed from an Azure Service Bus queue. Some messages contain malformed data and fail every time. After several attempts they stop appearing in the queue, and the team cannot find them. Where are the messages and what should you build?
- A. They are in the dead-letter queue after exceeding max delivery count; add a processor that reads and handles the dead-letter queue
- B. Service Bus deleted them; enable duplicate detection
- C. They were forwarded to a topic; add a subscription to the topic
- D. They were published to Event Grid; add an event subscription
Show answer
Answer: A
When a message exceeds the queue’s maximum delivery count, Service Bus moves it to the dead-letter subqueue. The fix is a process that reads the dead-letter queue, inspects the reason, and repairs or discards each message. Messages are not deleted automatically after failed processing, and neither topics nor Event Grid are involved in this failure path.
Want more questions like this? Full AI-200 practice tests →
Question 4. A Function app reads a database connection string. Security requires that the secret is never stored in code or app settings, that it can be rotated without redeploying, and that no credential is needed to reach it. What should you implement?
- A. Store the connection string as an environment variable in the container image
- B. Authenticate to Key Vault with a service principal client secret stored in app settings
- C. Store the connection string as a plain key-value in Azure App Configuration
- D. Give the Function app a managed identity with access to Key Vault, and use a Key Vault reference in the app setting
Show answer
Answer: D
A managed identity with a Key Vault role, combined with a Key Vault reference in the app setting, keeps the secret in Key Vault, needs no stored credential, and picks up a new version on rotation. An environment variable in the image bakes the secret into the artifact. A service principal secret is itself a stored credential. App Configuration without a Key Vault reference stores the value as plain configuration.
Want more questions like this? Full AI-200 practice tests →
Question 5. Requests to an API are traced with OpenTelemetry into Application Insights. You need a chart of the 95th percentile request duration in five-minute intervals over the last day. Which KQL query fits?
- A.
requests | summarize avg(duration) by bin(timestamp, 5m) - B.
requests | where timestamp > ago(1d) | summarize count() by bin(timestamp, 5m) - C.
requests | where timestamp > ago(1d) | summarize percentile(duration, 95) by bin(timestamp, 5m) | render timechart - D.
requests | where timestamp > ago(1d) | top 95 by duration
Show answer
Answer: C
Filtering on timestamp, summarising percentile(duration, 95) by bin(timestamp, 5m) and rendering a time chart produces exactly the requested series. Option A returns an average, not a percentile. Option B counts requests rather than measuring duration. Option D sorts raw rows and takes 95 of them, which is not a percentile at all.
Want more questions like this? Full AI-200 practice tests →
How did you do?
Questions 2 and 5 are the ones with little AZ-204 equivalent. If they slowed you down while the rest did not, you are where most former AZ-204 candidates are. Week 2 of the study plan covers the vector material, and the monitoring guide covers KQL.
For a longer check, take the 20-question practice test.