AI-200 Developing AI Cloud Solutions on Azure exam badge

Microsoft

Developing AI Cloud Solutions on Azure (AI-200) practice questions

★★★★★★ 4.7 69 reviews — aggregated from community votes by visitors who practiced this bank

Questions
69
Exam length
120 min
Passing score
700
  • Instant feedback per question
  • Explanation for every option
  • Free, no signup

Domain coverage 4 domains

  • Develop containerized solutions on Azure (20-25%)
  • Develop AI solutions by using Azure data management services (25-30%)
  • Connect to and consume Azure services (20-25%)
  • Secure, monitor, troubleshoot Azure solutions (20-25%)

Sample questions 7 of 69 shown

  1. Q1
    You maintain multiple versions of a container image in Azure Container Registry. The production deployment must always run the exact same image build even if tags are changed later. You need to ensure predictable and immutable image selection during deployment. What should you do?
    1. Tag the image as production and deploy it by using the production tag.
    2. Schedule nightly rebuilds of the image.
    3. Configure deployment to use the latest tag.
    4. Identify the image by using its SHA digest.
    Use the image's SHA-256 manifest digest when defining the production deployment. Azure Container Registry assigns every pushed image manifest a unique digest, and Microsoft explicitly states that pulling an image by digest guarantees the image version being retrieved, even if an identically named tag is later pushed to a different image. A digest reference has the form myregistry.azurecr.io/repository@sha256:<digest>. Tags such as production or latest are mutable references that can be overwritten, and a scheduled rebuild also creates a new artifact, so neither can guarantee exact reproducibility.
  2. Q2
    A container in an AKS cluster repeatedly restarts. Pod events show probe failures, although node-level CPU and memory metrics are normal. You need to diagnose the cause of the repeating restarts. What should you do first?
    1. Scale the deployment to more replicas.
    2. Decrease the initialDelaySeconds for the container liveness probe.
    3. Drain and reboot the node hosting the pod.
    4. Inspect the pod events and container logs.
    The first diagnostic step is to inspect the pod events and container logs. Because node-level metrics are normal, the cause is most likely at the pod or container level (for example a failing liveness/readiness probe, an application crash, or an OOMKill). Pod events and logs expose the probe failure reason and container exit details before you change any configuration. Scaling replicas, changing probe timing, or rebooting the node would act without evidence.
  3. Q3
    You develop a message-processing service deployed to Azure Container Apps. The service reads messages from an Azure Service Bus queue. The solution must minimize costs by ensuring NO compute resources are consumed when the queue is empty. You need to configure scaling for the service. Which two actions should you perform? Each correct answer presents part of the solution.
    1. Increase the scaling rule to allow for the maximum running replica count.
    2. Configure the scaling rule to allow for the termination of all active replicas.
    3. Configure a Kubernetes Event-driven Autoscaler rule that monitors queue length.
    4. Enable HTTP ingress concurrency scaling.
    Configure a KEDA scaling rule that monitors the Service Bus queue length (C) so replicas are driven by pending messages, and allow the scaling rule to scale the replica count down to zero (B) so no compute is consumed while the queue is empty. Increasing the maximum replica count does not reduce cost at idle, and HTTP concurrency scaling is unrelated to a queue-based workload.
  4. Q4
    You configure ACR Tasks to automate image builds. Container images must rebuild when application updates occur, when base image updates occur (such as when the underlying OS image is updated), and on regular scheduled rebuilds. You need to configure ACR Tasks to support automated image rebuilds. Which three triggers should you configure? Each correct answer presents part of the solution.
    1. Timer trigger
    2. Source code commit trigger
    3. Registry event trigger
    4. Base image update trigger
    5. Webhook notification trigger
    A source code commit trigger rebuilds the image when application code changes, a base image update trigger rebuilds when a dependent base image (for example the OS image) is updated, and a timer trigger performs the scheduled rebuilds. These three ACR Tasks triggers directly map to the three stated requirements.
  5. Q5
    You plan to deploy an Azure Container app. You need to configure the container app to support session affinity. Which ingress type and revision mode should you assign to the container app?
    1. TCP ingress type and single revision mode
    2. TCP ingress type and multiple revision mode
    3. HTTP ingress type and multiple revision mode
    4. HTTP ingress type and single revision mode
    HTTP ingress type and single revision mode must be assigned. Session affinity (sticky sessions) requires an HTTP-based proxy layer to track and route client requests to the same replica, which is incompatible with raw TCP ingress. Single revision mode ensures standard traffic flow without multi-revision splitting complexity.
  6. Q6
    Source records change before their cached copies expire, causing stale AI responses. The source database must remain authoritative. Which cache-aside change directly addresses the problem?
    1. Increase every cache entry's time to live.
    2. Treat the cached response as the system of record.
    3. Invalidate affected keys when source records change.
    4. Allow all clients to recompute a missing value concurrently.
    The source database remains authoritative, so a source update must make any corresponding cached representation unusable. Invalidating the affected keys at change time forces the next cache-aside read to miss and reload current source data. Increasing the Time to Live extends stale exposure, while treating the cache as authoritative reverses the required ownership.
  7. Q7
    A private endpoint was created for a data service, but an application still resolves the service name to its public address and public access remains enabled. What should be done?
    1. Configure private DNS and client routing, then explicitly restrict public access.
    2. Add outbound virtual network integration, then leave name resolution and public access unchanged.
    3. Recreate the private endpoint, then assume it disables every public network path.
    4. Require TLS on the public endpoint, then remove the application's private network route.
    Creating a private endpoint supplies a private network address, but clients use it only when DNS resolves the service name to that address and routing can reach it. Private DNS and client routing establish the private path. Public access is a separate setting, so it must be explicitly restricted after the private path is working to prevent unintended public connections.

Answer all 69 questions with instant feedback →

Reader reviews 69 ratings

4.7 ★★★★★★ Based on 69 ratings
  • Bought the AI-200 pack last month and honestly the detailed explanations are where this really shines.

    ★★★★★★ Ethan C.

  • I had almost given up on AI-200 after failing twice. These questions pinpointed exactly what I was missing. Third time’s the charm!

    ★★★★★ Noah S.

  • Solid AI-200 prep. No complaints. Questions are relevant and the platform works well on both desktop and phone.

    ★★★★★★ Scarlett W.

  • Solid Microsoft prep material. The AI-200 questions are well-written and the answer rationales are thorough.

    ★★★★★★ Madison T.

  • I bought access for the AI-200 exam as a gift for my brother. He passed on his first try and said the questions were spot-on.

    ★★★★★★ Cameron J.

  • I recommend this to everyone preparing for AI-200. The lifetime access is great — I keep coming back for reference.

    ★★★★★★ Robert C.

Log in to rate this exam and leave a review.

Frequently asked questions

Is AI-200 the replacement for AZ-204, and how is it different from AI-103?

Yes on the first count: AZ-204 (Developing Solutions for Microsoft Azure) retired on July 31, 2026, and AI-200 is the exam Microsoft points back-end developers to from that point on. The distinction with AI-103 matters too — AI-103 tests building the AI application itself (models, agents, retrieval in Microsoft Foundry), while AI-200 tests the cloud infrastructure beneath it: container hosting, vector-capable data services such as Cosmos DB for NoSQL and PostgreSQL with pgvector, Service Bus and Event Grid integration, and observability. The two are parallel Associate specializations, not prerequisites for each other. Our question bank stays strictly on the AI-200 side of that line.

What changed in the October 7, 2026 skills outline?

The refresh made Develop AI solutions by using Azure data management services the largest of the four domains at 25-30%, raising the share of vector-workload questions — Cosmos DB indexing policies and consistency levels, pgvector compute overhead, resource sizing for vector workloads, and Redis vector indexes. The outline also spells out RAG patterns with metadata filters under PostgreSQL and connection optimization for throughput. Older question sets written against the pre-October outline underweight this domain, so our practice mix follows the current weighting and flags the delta in explanations.

Which vector data services should I drill hardest for AI-200?

All three, and know what separates them. Cosmos DB for NoSQL questions revolve around Request Unit consumption — indexing policies, consistency levels, and the change feed processor. PostgreSQL questions revolve around pgvector: index strategies that cut query latency and compute overhead, sizing compute and memory for vector workloads, and RAG retrieval with metadata filters. Azure Managed Redis questions revolve around caching semantics — expiration and invalidation — plus HNSW vector indexes for similarity search. Expect scenarios that give you a latency or cost symptom and ask which store's which knob fixes it.

How long is the AI-200 exam, how many questions are there, and what score do I need?

The exam runs 120 minutes and requires a scaled score of 700 or greater out of 1000 — a scaled score, not a raw percent-correct. Microsoft does not publish a fixed item count for AI-200; its general guidance says role-based exams typically contain 40-60 questions, and community mock exams standardize on 50. Question formats include multiple choice, multi-select, and scenario items, and the exam may include interactive components. It is delivered through Pearson VUE online or at a test center, costs USD 165 in the US (regional pricing varies), and the certification renews annually through a free online assessment on Microsoft Learn.

Is there an AI-200 mock exam for timed, scenario-based practice?

Yes. Our AI-200 mock exam reproduces the decision-style items that dominate the blueprint — KEDA scale rules against a Storage queue, Event Grid retry behavior when an endpoint returns HTTP 400, choosing between a change feed processor and a pull-mode FeedIterator, or a pgvector index choice for a filtered similarity query blowing its latency SLA. Sitting the set under a 120-minute timer builds the pacing you need, and every answer explanation names the SDK call, index policy, or setting that settles the scenario.

Can I download the AI-200 questions as a PDF for offline study?

Absolutely. The full question bank ships as a downloadable PDF alongside the online practice questions and explanations, so you can re-read KEDA scale-rule parameters, Cosmos DB indexing policy JSON, pgvector index commands, Service Bus dead-letter configurations, and KQL query patterns during a commute or anywhere without stable internet — the same material, just portable and printable.

Where do Azure OpenAI, Microsoft Foundry, and Azure AI Search fit in AI-200?

Only as supporting context. The current outline has no top-level objective for provisioning models or agents — Azure OpenAI and Microsoft Foundry can appear as background in scenario framing (for example, a token-budget or deployment-rate-limit question), but the tested skills underneath are AI-200's own: App Configuration, Key Vault references, deployment-level rate limits, and the data-service plumbing. If a practice question is really about building agents or evaluating prompts in Foundry, it belongs to AI-103 and is deliberately absent here.

Free study resources

Community-verified analysis of 17 questions from real test-taker discussions — 4 deep analyses and 0 FAQs.