The short version
AI-103 has five domains. Plan and manage (25-30%) and generative AI and agents (30-35%) together make up most of the exam. The other three each carry 10-15%. This page follows the skills measured as of April 16, 2026, which Microsoft’s study guide says are “intended to illustrate” how skills are assessed, so related topics can appear too. The trap and practice notes below are our own general advice, not Microsoft’s statements.
Plan and manage an Azure AI solution (25-30%)
What it covers. Four areas. First, choosing the right model and Foundry service for a task: large language models, small language models, multimodal models, and the Foundry tools, plus methods for retrieval and indexing and services for memory, tools, and knowledge. Second, setting up solutions: designing Azure infrastructure, choosing deployment options, configuring model and agent deployments, and integrating Foundry projects with CI/CD. Third, managing, monitoring, and securing: quotas, scaling, rate limits and cost; model performance, drift, safety events, and grounding quality; data ingestion and search index health; and security with managed identity, private networking, keyless credentials, and role policies. Fourth, responsible AI: safety filters, guardrails, risk detection, evaluators, trace logging, provenance, approval workflows, and controls on agent behavior.
In practice. This is the “will it survive production” domain. A question might ask how an app should authenticate without storing a key, or what to change when requests hit rate limits.
Typical traps. Mixing up what a content filter does with what an evaluator does: one acts when a request runs, the other measures output afterward. Picking the most powerful model when the scenario asks about cost or latency. Forgetting that a security answer usually means identity and role assignment, not a shared secret.
What to practice. Deploy a model two ways and compare the settings. Call it with keyless authentication. Read a trace. Turn a safety filter on and see what it blocks.
Implement generative AI and agentic solutions (30-35%)
What it covers. Three areas. Building generative apps with Foundry: deploying and consuming models, retrieval-augmented generation, workflows and multistep reasoning pipelines, evaluating for fabrications, relevance, quality, and safety, using the Foundry SDKs, and connecting an app to a Foundry project. Building agents: defining roles, goals, conversation tracking, and tool schemas; integrating retrieval, function calling, and memory; connecting tools such as APIs, knowledge stores, search, Content Understanding, and custom functions; orchestrating multiple agents; building workflows with approval flow controls; and monitoring and error analysis. Optimizing: prompt engineering, model parameters, reflection and self-critique, observability with tracing and token analytics, and hybrid flows that mix models with rules engines.
In practice. You are the developer wiring a model to data and tools, and then proving it behaves.
Typical traps. Treating a prompt as a guarantee: a rule that must always hold belongs in code, not only in instructions. Confusing a tool’s schema (what the model sees) with the function itself (what your code runs). Assuming a single agent when the scenario describes specialist roles that need delegation. Forgetting that models do not remember earlier requests unless you pass the history or use a conversation object.
What to practice. Build a RAG app, then add an agent with a function tool and an approval step. Change one model parameter at a time and compare results. Run an evaluation and read the scores.
Implement computer vision solutions (10-15%)
What it covers. Image and video generation from text prompts and reference media, image editing such as inpainting and mask-based edits, video editing workflows, and the platform controls for generation. Multimodal understanding: captions, visual question answering grounded in what is in the image, alt text aligned to accessibility guidelines, Content Understanding for visual characteristics and video segments (single-task and pro-mode pipelines), and finding objects or regions. Responsible AI for visual content: unsafe-content filters, indirect prompt injection through text hidden in images, and visual policy rules such as watermarks and brand usage.
In practice. Think product photos, accessibility text, and moderation of user-uploaded images.
Typical traps. Remembering old image-model behavior when the question describes the current one. Overlooking that text inside an image can carry instructions to the model. Choosing a generation feature when the task is understanding, or the reverse.
What to practice. Generate an image, then edit part of it with a mask. Ask a multimodal model to caption and answer questions about a picture. Try an analyzer on a short video.
Implement text analysis solutions (10-15%)
What it covers. Language-model text analysis: extracting entities, topics, summaries, and structured JSON, detecting sentiment, tone, safety issues, and sensitive content, translating with Azure Translator in Foundry Tools or language-model flows, and tailoring outputs for domain tasks such as compliance summaries. Speech: speech to text and text to speech for agents, speech as an agent modality including custom speech models, reasoning over audio inputs, and translating speech.
In practice. Support-ticket triage, call summaries, multilingual content, and voice agents.
Typical traps. Picking a heavy language-model solution where a purpose-built service fits, or the opposite. Mixing up speech modes: live audio, a finished recording, and a large batch of stored files each need a different approach. Ignoring output format when the scenario requires structured data.
What to practice. Ask a model for JSON that matches a schema. Transcribe a short clip. Translate a paragraph two ways.
Implement information extraction solutions (10-15%)
What it covers. Retrieval and grounding pipelines: ingesting and indexing documents, images, audio, and video; configuring semantic, hybrid, and vector search; enrichment with built-in or custom skills; RAG ingestion with OCR; and connecting retrieval to workflows and agent tools. Document extraction: multimodal pipelines combining OCR, layout analysis, and field extraction, clean grounded outputs for agents and RAG with Content Understanding, and analyzers that produce structured or markdown output.
In practice. Turning a pile of PDFs, scans, and recordings into something an agent can search and cite.
Typical traps. Choosing vector search for exact identifiers like part numbers, where keyword matching is stronger. Skipping OCR for scanned pages. Forgetting that the order of enrichment steps matters.
What to practice. Index a small document set, run the same query with keyword, vector, and hybrid search, and compare. Process a scanned document with OCR and layout analysis.
More guides for AI-103
- AI-103 Exam Format, Scoring, and Retakes
- Is the AI-103 Exam Hard? An Honest Look
- AI-103 Prerequisites and Skills Self-Check
- AI-103 Renewal and Retakes Explained
- Is AI-103 Worth It? An Honest Look
- AI-103 Cheat Sheet: Key Facts to Review
- AI-103 vs AI-200: Which Exam Should You Take?
- AB-620 vs AI-103: Which Exam First?
Next step
Put the weightings to work in the study plan, or test yourself on the sample questions. The AI-103 practice tests give you more of the same by domain.