Azure AI-102 cheat sheet for key facts, traps, service mappings, and final review.
Use this cheat sheet for Microsoft Certified: Azure AI Engineer Associate (AI-102) after you know the basics but before you start a timed practice block. The goal is not to memorize a vendor catalog; the goal is to classify the scenario and reject attractive wrong answers quickly.
Use this when the stem mixes app planning, retrieval, multimodal capability, operations, or governance.
flowchart TD
S["Scenario"] --> W["Classify the workload"]
W --> B["Choose the Azure boundary and service family"]
B --> O["Separate build, secure, evaluate, or monitor"]
O --> V["Verify with telemetry, tests, and evidence"]
| Lane | Decision rule | Reject when |
|---|---|---|
| Foundry solution planning | Choose models, deployment shape, identity, network boundaries, cost controls, and responsible AI checks before coding. | Jumping straight to a model or prompt when the requirement is governance, private access, or deployment control. |
| Generative AI and agents | Separate prompt design, tool use, retrieval grounding, agent orchestration, evaluation, and safety filtering. | Treating every failure as prompt wording when the fix is retrieval quality, permissions, tool contract, or evaluation evidence. |
| Vision, speech, text, and extraction | Match the workload to the right multimodal, language, document, or content-understanding capability. | Using a general chat model when the scenario needs structured extraction, OCR, speech, or image-specific handling. |
| Python implementation and operations | Know where SDK calls, secrets, managed identity, telemetry, retries, and evaluation loops belong. | Shipping a demo client without auth, monitoring, error handling, or reproducible evaluation results. |
| Trap | Better instinct |
|---|---|
| Prompt-only answers for production problems | Ask whether the stem is really about data grounding, security, evaluation, deployment, or monitoring. |
| Ignoring managed identity and private networking | Secure app-to-service access before tuning model output. |
| Treating agent tools as magic | Check tool permissions, parameters, schema, failure behavior, and auditability. |
| No quality evidence | Prefer answers that evaluate groundedness, relevance, safety, latency, and regression behavior. |
| If the stem says | Start with |
|---|---|
| least privilege, private access, compliance, or audit | identity scope, data boundary, policy enforcement, logging, and ownership |
| least operational effort | managed service, native integration, simple workflow, and fewer moving parts |
| high availability, recovery, or outage | failure domain, recovery objective, health check, rollback, and validation |
| performance, scale, or cost | bottleneck evidence, traffic pattern, sizing, caching, batching, and quotas |
| troubleshoot, diagnose, or investigate | symptom, recent change, logs, metrics, status, dependency, and smallest safe test |
Use IT Mastery for the exact product route, practice status, spaced review when available, and close-answer explanation practice as coverage expands.
Open the exact IT Mastery route here: AI-102 on MasteryExamPrep.
Classify the workload first: agent, retrieval, multimodal, extraction, or operations. Then apply identity, data boundary, evaluation, and observability.