AI-103 · Associate
AI-103 Cheat Sheet 2026: Quick Reference
Every domain condensed to a table, plus a model-type decision guide and a Foundry agent tool-types reference for the exam's most-confused pairs.
Exam Snapshot
Questions: ~40
Duration: 120 minutes
Passing score: 700 / 1000
Cost: $165 USD
Level: Associate
Last updated: 2026-04-16
Every domain weight and reference link below is pulled directly from Microsoft's official AI-103 exam page and cross-checked against our 500-question AI-103 bank, which we audit against Microsoft Learn documentation before anything goes live.
1. Plan and Manage an Azure AI Solution (25-30%)
| Concept | Area | Quick note |
|---|---|---|
| Choose model type | Model selection | LLM, small language model, multimodal model, or Foundry Tools — match to the task |
| Choose Foundry service | Service selection | Generative, grounding, vector search, agent workflows, or multimodal processing |
| Retrieval & indexing method | Service selection | Choosing how content gets found later, not just stored |
| Memory, tool, knowledge integration | Agent design | What an agent can remember, call, or query |
| Infra design for AI apps/agents | Setup | Architecture decisions before anything is deployed |
| Deployment options & config | Setup | How a model or agent actually goes live |
| CI/CD integration with Foundry projects | Setup | Automated promotion, not manual deploys |
| Quotas, scaling, rate limits, cost | Manage | Keeping workloads within budget and limits |
| Monitor performance, drift, safety, grounding | Manage | Ongoing model health, not a one-time check |
| Monitor ingestion, index health, relevance | Manage | RAG pipeline health specifically |
| Managed identity, private networking, keyless, RBAC | Security | No passwords, no public exposure |
| Safety filters, guardrails, content moderation | Responsible AI | Blocking unsafe output before it ships |
| Trace logging, provenance, approval workflows | Responsible AI | Auditability of what the system did and why |
| Oversight modes, constraints, tool-access controls | Responsible AI | Governing how much independence an agent gets |
2. Implement Generative AI & Agentic Solutions (30-35%)
| Concept | Area | Quick note |
|---|---|---|
| Deploy/consume model types | Generative apps | LLM, small, code, or multimodal models |
| RAG implementation | Generative apps | Grounding responses in your own data |
| Workflows & multistep reasoning | Generative apps | Chaining steps rather than one call |
| Evaluate models/apps | Generative apps | Fabrications, relevance, quality, safety — before shipping |
| Foundry SDKs & connectors | Generative apps | The integration layer between app and Foundry |
| Connect app to a Foundry project | Generative apps | The base wiring step everything else depends on |
| Define agent roles, goals, tool schemas | Agents | Where agent design starts |
| Integrate retrieval, function-calling, memory | Agents | Core agent capabilities |
| Integrate agent tools | Agents | APIs, knowledge stores, search, Content Understanding, custom functions |
| Orchestrated multi-agent solutions | Agents | Coordinating more than one agent |
| Autonomous/semiautonomous with safeguards | Agents | How much independence an agent is given |
| Monitor deployed agents, error analysis | Agents | Ongoing agent health, not just launch-day |
| Prompt engineering & model parameters | Optimize | Tuning generation behavior directly |
| Model reflection, chain-of-thought, self-critique | Optimize | Self-correcting techniques |
| Observability | Optimize | Tracing, token analytics, safety signals, latency |
| Orchestrate multiple models/flows | Optimize | Combining approaches, including hybrid rules + LLM |
3. Implement Computer Vision Solutions (10-15%)
| Concept | Area | Quick note |
|---|---|---|
| Text-to-image generation | Image/video gen | Generate from text prompts and reference media |
| Text-to-video generation | Image/video gen | Same idea, video output |
| Image editing | Image/video gen | Inpainting, mask-based, prompt-driven edits |
| Video editing | Image/video gen | Modifying already-generated video |
| Visual context analysis | Understanding | What a multimodal model sees in an image |
| Captions (single/multi-image) | Understanding | Concise or detailed descriptions |
| Visual question-answering | Understanding | Q&A grounded in visual evidence |
| Alt-text & accessibility descriptions | Understanding | Accessibility-specific output, not just a caption |
| Content Understanding (visual) | Understanding | Structured extraction of visual characteristics |
| Video analysis workflows | Understanding | Processing and interpreting video segments |
| Unsafe visual content filters | Responsible AI | Blocking disallowed imagery |
| Indirect prompt injection via image text | Responsible AI | A vision-specific attack vector, not just text prompts |
| Visual policy rules | Responsible AI | Watermarks, brand rules, prohibited symbols |
4. Implement Text Analysis Solutions (10-15%)
| Concept | Area | Quick note |
|---|---|---|
| Entity/topic/summary/JSON extraction | Text analysis | Generative prompting plus Foundry Tools |
| Sentiment, tone, safety, sensitive content | Text analysis | Classification-style detection tasks |
| Translation | Text analysis | Azure Translator or LLM-powered — two distinct approaches |
| Domain-specific output customization | Text analysis | e.g. compliance summarization |
| Speech-to-text / text-to-speech | Speech | For agentic interactions specifically |
| Speech as an agent modality | Speech | Including custom speech models |
| Multimodal reasoning from audio | Speech | Beyond simple transcription |
| Speech translation | Speech | Translating spoken input directly |
5. Implement Information Extraction Solutions (10-15%)
| Concept | Area | Quick note |
|---|---|---|
| Ingest & index docs/images/audio/video | Retrieval pipelines | Where the RAG data pipeline starts |
| Semantic, hybrid, vector search | Retrieval pipelines | Three distinct grounding search types |
| Enrichment via custom/built-in skills | Retrieval pipelines | Adding structure during ingestion |
| RAG ingestion with OCR | Retrieval pipelines | For scanned or image-based documents |
| Connect retrieval to agent tools | Retrieval pipelines | Making retrieval usable by an agent, not just a search box |
| Multimodal extraction | Document extraction | OCR plus layout analysis plus field extraction |
| Content Understanding (grounded output) | Document extraction | Clean output ready for RAG or agents |
| Structured/markdown analyzer output | Document extraction | Format built for downstream reasoning |
Model Type Quick Decision
This choice shows up across Domains 1 and 2, not just once:
| Need | Use this |
|---|---|
| General-purpose reasoning, broad knowledge | Large language model (LLM) |
| Narrow task, low latency, cost-sensitive, or constrained environment | Small language model |
| Input includes images, audio, or video alongside text | Multimodal model |
| A managed capability rather than a raw model (search, Content Understanding) | Foundry Tools |
Foundry Agent Tool Types Quick Reference
| Tool type | Use for |
|---|---|
| File search | Retrieving answers grounded in uploaded documents |
| Code interpreter | Running code to compute, analyze, or transform data |
| Function calling | Calling your own API or backend function |
| Knowledge store | Persistent, structured knowledge the agent can query |
| Content Understanding | Extracting structured data from multimodal sources |
| Custom function | Anything not covered by a built-in tool |
AI-103 FAQ
How is this cheat sheet different from the AI-103 study guide?
The study guide explains the most-tested topics in narrative form with a 4-week plan. This page is tables only, for a final scan before the exam.
When should I use a small language model instead of an LLM?
When the task is narrow, latency-sensitive, cost-sensitive, or needs to run in a constrained environment — an LLM is the default for general-purpose reasoning otherwise.
What's the fastest way to review right before the exam?
Scan the five domain tables plus the model-type decision guide and agent tool-types reference — both distinctions come up across multiple domains.
Can I print this AI-103 cheat sheet?
Yes — every table here is plain HTML, so a browser print or "print to PDF" renders cleanly.
MSCertQuiz sells practice-exam access for AI-103 and other Microsoft certifications; this cheat sheet is written by the same team that builds those questions.
Keep Studying
AI-103 Study Guide
The 4-week plan behind this quick reference.
Free AI-103 Practice Questions
Scenario questions with detailed explanations.
AI-103 vs. AI-200
How these two related exams differ.
AI-103 Certification Overview
Full exam details and the complete 500-question practice bank.
Reviewed the cheat sheet? Now drill it.
Start with free AI-103 questions covering every domain above.
Start Free AI-103 Practice