I Studied for AI-200 for 3 Weeks — Here Is What Actually Appeared

My starting point

I held AZ-204 (Developing Solutions for Microsoft Azure) and worked daily with AKS, Azure Functions, and Service Bus. That gave me roughly 50% of AI-200 for free. The new material for me was the AI data layer: Cosmos DB vector search, pgvector in PostgreSQL, Redis vector capabilities, and the operational mindset around KQL queries and monitoring AI workloads.

If you do not have an AZ-204 background, add at least 2 more weeks and go deeper on the containers domain.

Week 1 — Plug the gaps in containers

Even though I knew AKS day-to-day, I needed to be exam-precise. The exam tests concepts you use but never need to articulate:

• AKS network plugins — kubenet vs Azure CNI vs CNI Overlay. Know when you need which, and how many IPs each consumes.
• Workload Identity — the OIDC federation model, how it replaces pod-managed identity, and what EnableOIDCIssuer does.
• Container Apps scaling rules — HTTP, CPU, and custom KEDA triggers. Know the difference between min-replicas 0 and 1.
• ACR tasks — base image triggers, multi-step YAML tasks, az acr import. More questions on ACR than I expected.

I read every module page on this site and did the sample questions. Anything I got wrong, I re-read that section.

Week 2 — The AI data layer (the new material)

This is where AI-200 differs from AZ-204 and where most candidates underinvest. Three areas dominated my study time:

Cosmos DB vector search

The exam tested DiskANN specifically — what it is (disk-based ANN algorithm), how to configure the vector policy (dimensions, metric, index type), and when to use cosine vs Euclidean distance. Know that cosine is standard for text embeddings. Know that cross-partition queries are expensive. Know TTL and change feed — both appeared.

PostgreSQL + pgvector

More questions here than I expected. Know the two index types (IVFFlat and HNSW), when to use each, and how to tune them (ivfflat.probes, hnsw.ef_search). Know the distance operators: <-> (L2), <=> (cosine), <#> (inner product). Flexible Server HA modes (same-zone vs zone-redundant) appeared at least twice.

Redis for AI

Know the tier differences (Basic/Standard/Premium/Enterprise). Know that vector search requires Enterprise tier with RediSearch. Sliding window rate limiting with sorted sets appeared in a scenario question.

Week 3 — Messaging, security, and drilling questions

I spent the first two days on the messaging layer (Service Bus sessions, dead-letter queues, Event Grid retry policy, Event Hubs consumer groups). Then I shifted entirely to doing practice questions — at least 50 per day. Patterns I noticed:

• Scenario questions ask you to choose between two correct-sounding options. The differentiator is always a specific constraint in the scenario (no internet, must be atomic, must scale to zero).
• KQL appeared in at least 3 questions. Know: | where, | summarize, ago(1h), | top N by.
• Managed identity vs service principal — always prefer managed identity unless the question specifies a cross-tenant or non-Azure scenario.
• RBAC: know the difference between Owner (can assign roles), Contributor (cannot assign roles), and least-privilege patterns.

What actually appeared on exam day

I cannot share specific questions. What I can say about the domain distribution (my experience — yours may differ):

• ~25% — Containers (AKS, ACR, Container Apps). Heavier on AKS than I expected.
• ~25% — AI data services (Cosmos DB vector, pgvector, Redis). This was the hardest section for me.
• ~20% — Messaging and event-driven (Service Bus sessions, Event Grid retry, Event Hubs partitions).
• ~30% — Security and operations (Key Vault, managed identity, RBAC, KQL, Container Insights).

The exam is 60 questions in 100 minutes. I finished in 70 minutes and had time to review flagged questions. Several questions had a case study format with a shared scenario. Read the scenario thoroughly before the questions — the constraint that changes the answer is usually buried in the scenario text.

What I would do differently

• Spend more time on pgvector index tuning — I underestimated how many questions it generated.
• Practice writing KQL queries by hand, not just recognising them.
• Do the mini-mock exam under timed conditions at least 3 times in week 3.
• Use the official Microsoft Learn paths — they are dry but they match the question language exactly.