AI-103 Cheat Sheet: Key Facts to Review

The short version

This sheet is a last-pass review for AI-103, Developing AI Apps and Agents on Azure. It helps you recall and compare ideas; it does not cover everything on the exam, and it is no substitute for hands-on work. It follows Microsoft’s skills measured as of April 16, 2026. Verify details against the Microsoft study guide before exam day.

Exam facts at a glance

Item	Fact (Microsoft, checked October 22, 2026)
Exam name	Developing AI Apps and Agents on Azure
Level	Intermediate, associate
Time	120 minutes
Pass mark	700 on a scale of 1 to 1,000
Delivery	Proctored, scheduled through Pearson VUE; may include interactive components
Questions	Number not published
Retake	24 hours, then 14 days; five attempts in 12 months
Credential validity	One year, free renewal assessment

Domain weights: plan and manage 25-30%, generative AI and agents 30-35%, computer vision 10-15%, text analysis 10-15%, information extraction 10-15%.

Domain 1: plan and manage (25-30%)

Key concepts

Match the model to the task: large, small, or multimodal model, judged by quality, cost, and latency.
Deployment choices, quotas, rate limits, and scaling.
Security: managed identity, keyless credentials, private networking, role assignments.
Responsible AI: filters, guardrails, evaluators, tracing, approval workflows.

Pairs that get confused

Content filter acts on a request or response while the app runs. Evaluator measures quality or safety after the fact.
Quota is how much capacity you may use. Rate limit is how fast requests are accepted.
Monitoring tells you something changed. Tracing shows the steps of one request.

Decision rules

Scenario says “no secrets in code” or “no stored key”: think managed identity and role assignment.
Scenario says “traffic must not use the public internet”: think private networking.
Scenario stresses cost or speed on a simple task: think a smaller model.

Domain 2: generative AI and agents (30-35%)

Key concepts

Retrieval-augmented generation: retrieve relevant content, pass it to the model, answer from it.
Agents: role, goal, tools with schemas, memory or conversation tracking, orchestration of multiple agents.
Optimization: prompt design, model parameters, reflection and self-critique, tracing and token analytics.

Pairs that get confused

Tool schema is what the model sees. Function is the code your app runs.
Prompt rule is a request. Code check or approval step is enforcement.
Single agent suits one role. Multi-agent suits separate specialist roles with delegation.
Conversation state must be supplied or stored; the model does not remember earlier calls by itself.

Decision rules

“Must always” or “must never”: enforce in code or a safeguard, not only in the prompt.
“Answers must come from company documents”: retrieval and grounding.
“A human must confirm before action”: an approval flow.
“Output varies too much”: lower randomness settings before rewriting everything.

Domain 3: computer vision (10-15%)

Key concepts

Generating and editing images and video from prompts and reference media; inpainting uses a mask to change one region.
Multimodal understanding: captions, questions about an image, alt text, object or region finding, video segments.
Safety for visual content, including text hidden in an image that tries to instruct the model.

Pairs that get confused

Generation creates new media. Understanding describes or analyzes existing media.
Caption is short. Extended description and alt text serve accessibility needs.

Decision rules

“Change only part of an image”: mask-based editing.
“User-uploaded images may contain instructions”: treat embedded text as untrusted input.

Domain 4: text analysis (10-15%)

Key concepts

Extracting entities, topics, summaries, and structured JSON; detecting sentiment, tone, and sensitive content; translation.
Speech to text, text to speech, custom speech, and reasoning over audio.

Pairs that get confused

Live audio, a finished recording, and a large batch of stored files call for different approaches.
Purpose-built service versus language model prompt: use the simpler tool when it fits the task.

Decision rules

“Downstream code needs a fixed format”: ask for structured output that matches a schema.
“Voice conversation with an agent”: think speech in, speech out.

Domain 5: information extraction (10-15%)

Key concepts

Ingest and index documents, images, audio, and video; keyword, semantic, hybrid, and vector search; enrichment skills.
OCR and layout analysis for scans; Content Understanding analyzers that produce structured or markdown output.

Pairs that get confused

Keyword search matches exact terms. Vector search matches meaning. Hybrid combines both.
OCR reads text from images. Layout analysis finds structure such as tables and sections.

Decision rules

Exact identifiers such as part numbers: keyword matching is strong.
Scanned pages: OCR comes before indexing.
Mixed queries: hybrid search is often the safe choice.

Common traps

Picking the biggest model when the scenario asks for low cost.
Treating a prompt instruction as a firm rule.
Choosing generation when the task is understanding, or the reverse.
Reading only the first sentence of a scenario and missing a constraint such as “without keys” or “must stay private”.
Spending too long on one hard item. Microsoft says there is no penalty for guessing, so answer everything.

Last-day checklist

Confirm your appointment, ID, and the online or test center rules with Pearson VUE.
Skim this sheet and the domains explained page once, not five times.
Do not start a new topic.
Remember the break rule: after you start a break, you cannot go back to earlier questions.
Sleep. Eat. Arrive or log in early.

More guides for AI-103

Next step

Test recall with the sample questions and review the format on the exam format page. For more practice by domain, try the AI-103 practice tests. Back to the AI-103 overview.

Try the AI-103 practice exams