Browse Microsoft Certification Guides

Azure AI-102 Cheat Sheet

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.

First-pass question triage

  1. Name the tested lane before reading the answer choices.
  2. Underline the constraint: security, cost, reliability, latency, governance, implementation effort, or evidence.
  3. Reject answers that solve a neighboring problem but not the stated requirement.
  4. Prefer the smallest correct control, service, workflow, or command that satisfies the constraint.
  5. Look for proof: logs, tests, metrics, policy evidence, deployment status, evaluation results, or user-visible recovery.

AI-102 answer sequence

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"]

What to know cold

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.

Common traps and better instincts

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.

Final 15-minute review

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

Practice fit

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.

Decision order

Classify the workload first: agent, retrieval, multimodal, extraction, or operations. Then apply identity, data boundary, evaluation, and observability.

Revised on Sunday, May 10, 2026