Use this cheat sheet for Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) when you need to decide quickly between app architecture, model workflow, agent behavior, Azure AI service fit, and operational controls. This page intentionally avoids exact live exam facts; verify current Microsoft Learn status before final scheduling.
Read every scenario in this order
- Name the workload: app/agent planning, generative AI, retrieval, vision, speech, language, extraction, evaluation, or operations.
- Identify the Azure boundary: identity, network, data store, model deployment, service endpoint, or tool permission.
- Decide whether the answer is about building, securing, evaluating, or monitoring.
- Prefer the simplest managed path that satisfies the stated constraint.
- Reject answers that improve a demo but skip production proof: auth, telemetry, evaluation, and data governance.
AI-103 answer sequence
Use this when the stem mixes app choice, model fit, retrieval, tools, governance, and monitoring.
flowchart TD
S["Scenario"] --> W["What is the workload trying to do?"]
W --> F["Pick app, RAG, agent, extraction, or safety path"]
F --> G["Check identity, data boundary, and service fit"]
G --> V["Validate with telemetry, evaluation, and rollback"]
Foundry and app planning chooser
| Requirement |
Strong first answer |
| choose or deploy a model |
Match model capability, latency, cost, region, content policy, and deployment constraints. |
| build a chat app over enterprise data |
Use RAG with governed retrieval, source filtering, prompt grounding, and evaluation. |
| build an agentic workflow |
Define instructions, tools, permissions, state, error handling, and traceability. |
| connect from code securely |
Use managed identity or approved credential flow, not secrets embedded in code. |
| control production rollout |
Use versioned prompts/config, tests, monitoring, approval, and rollback. |
| prove quality |
Use evaluation data for groundedness, relevance, safety, latency, and regression behavior. |
Agent decision rules
| If the agent must… |
Check this before choosing an answer |
| call a business system |
tool contract, input validation, authorization, audit log, and error behavior |
| retrieve documents |
source permissions, index freshness, chunking, ranking, metadata, and citation needs |
| take an action |
human approval, idempotency, rollback, and least-privilege access |
| process user files |
file type, extraction service fit, malware or content checks, and storage boundary |
| handle sensitive data |
redaction, retention, encryption, private access, and logging restrictions |
Azure AI service fit
| Workload clue |
Better starting point |
| classify, summarize, translate, or understand text |
language or generative AI service depending on output flexibility and determinism |
| extract fields from forms or documents |
document extraction/content understanding path with schema validation |
| analyze images or video frames |
vision or multimodal model path, depending on task and output shape |
| speech-to-text or text-to-speech |
speech service path with language, voice, latency, and accessibility constraints |
| moderate unsafe content |
content safety and policy controls, plus workflow ownership |
| search enterprise content semantically |
search/vector index plus retrieval filters and source governance |
RAG quality checklist
| Quality problem |
Stronger fix |
| answer is plausible but unsupported |
improve grounding, retrieval ranking, citation handling, and evaluation |
| answer uses old information |
refresh source ingestion and index update process |
| answer sees forbidden content |
enforce identity-aware filtering and source permissions |
| answer misses key facts |
improve chunking, metadata, query rewriting, and prompt instructions |
| response is too slow |
reduce context, optimize retrieval, choose a suitable model, and measure each hop |
| behavior regresses after a change |
run repeatable evaluations before promotion |
Python implementation reminders
| Code-path issue |
Exam instinct |
| credentials |
Prefer managed identity or secure secret handling. Never hard-code keys in app code. |
| retries |
Retry transient failures with backoff only when the operation is safe to retry. |
| tool calls |
Validate inputs and outputs; log failures without leaking secrets. |
| telemetry |
Emit enough logs, traces, metrics, and correlation IDs to diagnose the request path. |
| config |
Keep endpoints, model deployments, prompts, and thresholds versionable and reviewable. |
| evaluation |
Treat evaluation as part of CI or release gating, not only as a manual spot check. |
Security, privacy, and responsible AI
| Concern |
What the exam usually rewards |
| least privilege |
managed identity, scoped roles, private endpoints where required, and key access review |
| sensitive input |
minimize retention, redact where appropriate, avoid logging prompt secrets, and classify data |
| unsafe output |
content filtering, policy, evaluation, human review, and escalation path |
| prompt injection |
restrict tool permissions, isolate instructions from retrieved content, validate tool arguments |
| audit |
preserve request IDs, model/config versions, tool actions, and approval evidence |
| compliance |
connect technical controls to data boundary, ownership, retention, and monitoring |
Common traps
| Trap |
Better instinct |
| Treating every failure as prompt wording |
Check retrieval, identity, tool contract, model fit, and evaluation evidence first. |
| Choosing a generic chat model for extraction |
If the output needs structured fields, validation, or repeatability, use an extraction-aware path. |
| Letting agents call broad tools |
Agents need stricter permissions and audit than normal app code, not looser ones. |
| Shipping without monitoring |
Production AI needs telemetry for latency, errors, safety, cost, and quality. |
| Ignoring data boundary |
Prompts, retrieved chunks, embeddings, logs, and outputs all need governance. |
Final 15-minute review
| If the stem says… |
Start here |
| app or solution planning |
Foundry project/deployment choice, identity, network, data, evaluation, and cost |
| agent |
instructions, tools, permissions, state, approval, and traceability |
| enterprise knowledge |
retrieval, indexing, metadata, source permissions, and grounded output |
| document/image/speech/text processing |
pick the specialized Azure AI capability before reaching for generic generation |
| responsible AI |
safety policy, evaluation, monitoring, human oversight, and privacy |
| Python implementation |
SDK flow, credentials, retries, telemetry, config, and release gates |
Practice fit
Use IT Mastery for the exact product route, practice status, spaced review when available, and close-answer explanation practice as coverage expands.
One-line decision rule
AI-103 answers should look like production Azure AI engineering: correct service fit, scoped identity, governed data, repeatable evaluation, safe agent tools, and observable runtime behavior.