code
AI-200
level
Associate
pass
700 / 1000
cost
$165 USD
questions
~40–60
duration
100 min
renewal
free / annual
track
Azure AI Cloud Dev
AI-200 has four domains and only one of them breaks 25% — data management. That single fact should drive your review order: if your time is short, the data-management domain (Cosmos DB, PostgreSQL/pgvector, Azure Managed Redis) is worth more per hour of study than any other section on this page.
This domain covers how you package, host, and scale the containers that run your AI workload — from building the image to keeping the running app healthy under unpredictable load. Expect scenario questions that describe a hosting constraint (cost, orchestration complexity, autoscaling behavior) and ask you to pick the Azure service or configuration that satisfies it.
ACR
Azure Container Registry — build, store, version images. Pairs with ACR Tasks for auto-rebuild on Git commit or base-image update.
App Service (containers)
Deploy a single container with App Settings / Key Vault references for env vars and secrets — no orchestration.
Container Apps revisions
Multiple revisions run side by side; traffic-split weights enable canary rollouts.
KEDA
Kubernetes Event-Driven Autoscaling built into Container Apps — scale rules on queue length, topic depth, etc., including scale-to-zero.
AKS manifests
Deploy/manage via YAML manifests; troubleshoot CrashLoopBackOff via readiness/liveness probes, pod events, and logs.
$ most-tested: fixed replica counts vs. KEDA scale rules for bursty, unpredictable load — KEDA wins whenever the scenario says “unpredictable” or “scale to zero.”
The exam repeatedly tests whether you can pick the right data store for a vector-search or general data-access scenario, not just whether you know each service exists. Read every scenario for its actual constraint — consistency requirement, existing schema shape, or cache-vs-source-of-truth role — because two of the three stores below will usually look technically capable of the same task.
Cosmos DB — RU tuning
Exclude unqueried paths from the indexing policy to cut write RU cost; choose consistency level (Session for monotonic-per-client, Strong for cross-region correctness) by requirement.
Cosmos DB — vector search
Requires a vector embedding policy + vector index on the container; query with the native vector distance function.
Cosmos DB — change feed
Change feed processor + lease container reacts to inserts/updates in near real time, no polling.
pgvector
Vector column on Azure Database for PostgreSQL; HNSW/IVFFlat approximate index trades a little recall for much lower latency at scale.
pgvector RAG
Vector similarity search combined with metadata filters (WHERE clauses) for retrieval-augmented generation over relational data.
Azure Managed Redis
Cache layer: caching + expiration + invalidation, plus vector indexing — sits in front of a durable store, is not the source of truth.
$ most-tested: which store wins a vector-search scenario. Cosmos DB → change-feed reactivity + NoSQL flexibility. PostgreSQL/pgvector → already-relational data + metadata-filtered RAG. Redis → hot-query cache in front of a slower source of truth.
This domain is about wiring services together reliably — queueing work, reacting to events, and exposing serverless endpoints — without losing messages or coupling systems too tightly. The recurring theme is durability and independence: can a consumer fail temporarily without losing data, and can new consumers be added without touching existing ones?
Service Bus dead-letter
Messages exceeding MaxDeliveryCount auto-move to the dead-letter sub-queue — inspect/reprocess without blocking the main queue.
Service Bus topics/subs
Pub/sub with independent, durable per-consumer subscriptions — each can lag/reprocess independently.
Event Grid filters
Declarative subject / event-type / advanced filters on the subscription — no application-code filtering needed.
Azure Functions bindings
Trigger (e.g. Service Bus) + output binding (e.g. Cosmos DB) — declarative wiring, minimal boilerplate.
$ most-tested: Service Bus topics (durable, per-consumer, reprocessable) vs. Event Grid (lightweight, fire-and-forget, filtered).
The last domain checks two related skills: locking down secrets and configuration with least privilege, and being able to actually find the cause of a production problem afterward. Expect questions where the "obvious" broad-permission answer is wrong precisely because it violates least privilege.
Key Vault rotation
Rotation policy on the secret + unpinned version reference in the app = auto-rotation with no redeploy.
Key Vault least privilege
Scope RBAC/access policy to Get only for runtime secret retrieval — not Administrator, not Owner.
App Configuration
Central store for non-secret settings; holds Key Vault references for secrets so the app reads from one place.
OpenTelemetry
Instrument each hop (Container Apps → Function → Cosmos DB); correlate spans by trace ID to find the slow hop.
KQL
where filters, summarize count() by groups, top N by ranks — the core pattern for log/metric queries in Log Analytics.
$ test --readiness
Try 40 Free AI-200 Practice Questions
Scenario-based, calibrated harder than the real exam. No card required.
Start Free Practice →A short, practical list for the morning of the exam — not more material to learn, just the order to review it in when time is limited.
- Re-read the four “most-tested” callouts above — they cover the distinctions that decide close-call questions, in the domains most likely to appear.
- Confirm you can explain, out loud, why Cosmos DB, pgvector, and Azure Managed Redis are each the right (or wrong) fit for a given vector-search scenario — the single highest-value comparison on the exam.
- Skim the acronym list once — recognizing HNSW, KEDA, and RU by sight saves reading time on scenario questions that assume you already know them.
- Do not cram new material the night before. AI-200 rewards calm scenario analysis over freshly memorized facts.
# Is this cheat sheet enough to pass AI-200 on its own?
No. AI-200 is scenario-heavy and assumes hands-on backend engineering experience. Use this as a final-review consolidation, paired with hands-on time in a free Azure trial subscription and full-length practice questions — not as a substitute for either.
# What is the single biggest domain to prioritize?
Develop AI solutions by using Azure data management services (25–30%) — Cosmos DB, PostgreSQL/pgvector, and Azure Managed Redis. It is the only domain above 25%, and the vector-search comparison across all three stores is the most-tested distinction on the exam.
# What is the passing score and cost for AI-200?
The passing score is 700 out of 1000. The exam costs $165 USD, runs about 100 minutes, and has roughly 40–60 questions.
# Cosmos DB, pgvector, or Azure Managed Redis — how do I remember which is which?
Cosmos DB when you need change-feed-driven reactivity and flexible NoSQL documents. PostgreSQL/pgvector when your data is already relational and you need RAG with metadata filters. Azure Managed Redis when the vectors are a cache in front of a slower source of truth — not the source of truth itself.
# Is this cheat sheet current with the latest AI-200 exam objectives?
Yes. It is built directly from Microsoft's official AI-200 skills-measured outline (last updated 2026-04-15), the same source used in our full AI-200 study guide, so the domain weights and sub-skills here match what Microsoft Learn currently publishes.
# How does a cheat sheet differ from the full AI-200 study guide?
The study guide walks through each domain in depth with a week-by-week plan for building the underlying skills. This cheat sheet assumes you've already studied and is meant for rapid final review the week of your exam — short definitions and the specific distinctions that decide close-call questions, not step-by-step instructions.
# What should I review last, the night before the exam?
The "most-tested" callouts under each domain below. They capture the specific decision points (which data store, which scaling rule, which messaging service) that repeatedly separate correct answers from plausible-looking wrong ones — a higher-yield use of your last hour than re-reading full documentation.
about.md
MSCertQuiz provides 500 practice questions per certification, calibrated harder than the real exam so test day feels easier. Start with 40 free questions on AI-200 — no card required.