AI-103 Developing AI Apps and Agents on Azure

700 AI-103 practice test questions on Microsoft Foundry, agents, RAG and vision. Explanations that teach the topic, never the answer. 70% to pass.

700 practice questions for AI-103 Developing AI Apps and Agents on Azure, covering all five exam domains at their published weightings, with explanations written to teach the topic rather than restate the answer.

700 AI-103 practice test questions for Microsoft Certified: Azure AI Apps and Agents Developer Associate, written to the skills measured as of 16 April 2026. AI-103 is the successor to AI-102, which retired on 30 June 2026, and it is a different paper rather than a renamed one: it is built around Microsoft Foundry, agents and grounding. AI-102 tested agents in a group worth 5 to 10 percent; on AI-103 they anchor the largest domain on the paper. Every question is tagged with the objective it tests, and the bank is weighed to match Microsoft's published domain split.

What is covered:

  1. Plan and manage an Azure AI solution - 196 questions. Choosing a model for the work in front of you, across large language models, small language models, multimodal models and Foundry Tools, and choosing the Foundry services for generation, grounding, vector search and agent workflows. Retrieval and indexing methods, and the memory, tool and knowledge integration an agent need. Designing the Azure infrastructure, choosing deployment options, configuring model and agent deployments, and putting Foundry projects into a CI/CD pipeline. Quotas, scaling, rate limits and cost. Monitoring model performance, drift, safety events and grounding quality, and monitoring ingestion quality and search index health. Security with managed identity, private networking, keyless credentials and role policies. Responsible AI in practice: safety filters and guardrails, evaluators and safety evaluations, trace logging and provenance metadata, approval workflows, and governing agent behavior with oversight modes and tool-access controls.
  2. Implement generative AI and agentic solutions - 224 questions. Deploying and consuming language, code and multimodal models. Retrieval-augmented generation in a real application, tool-augmented flows and multistep reasoning pipelines. Evaluating models and apps for fabrications, relevance, quality and safety. Foundry SDKs, connectors and the Responses API, and connecting an application to a Foundry project. Agents in depth: roles, goals, conversation tracking and tool schemas, agents that combine retrieval with function calling and memory, and agent tools including APIs, knowledge stores, search, MCP servers and custom functions. Orchestrated multi-agent solutions with the Microsoft Agent Framework, and autonomous or semiautonomous workflows with safeguards and approval controls. Prompt engineering and parameter tuning, reflection and self-critique loops, and observability through tracing, token analytics, safety signals and latency breakdowns.
  3. Implement computer vision solutions - 98 questions. Generating images from text prompts and reference media, generating and editing video, and image editing through inpainting, mask-based edits and prompt-driven modification, with the generation and editing controls the platform provides. Analyzing visual context with multimodal models, producing concise and detailed captions, answering questions grounded in visual evidence, and generating alt-text and extended descriptions to accessibility guidelines. Content Understanding for visual characteristics, including single-task and pro-mode pipelines, and video analysis that identifies objects, components and regions. Responsible AI for multimodal content: filtering unsafe visual material, enforcing visual policy rules, and detecting indirect prompt injection carried in text embedded in an image.
  4. Implement text analysis solutions - 91 questions. Extracting entities, topics, summaries and structured JSON output with generative prompting, detecting sentiment, tone, safety issues and sensitive content, translating with Azure Translator and with model-powered flows, and customizing outputs for domain work such as compliance summarization. Speech is in this domain and the domain name does not say so, which is how people miss it: 44 of these 91 questions are speech to text and text to speech for agent interactions, speech as an agent modality including custom speech models and Voice Live, multimodal reasoning from audio input, and translating speech into other languages.
  5. Implement information extraction solutions - 91 questions. Ingesting and indexing documents, images, audio and video. Semantic, hybrid and vector search for grounding, and the ranking behavior that decides what a model actually sees. Enrichment with built-in and custom skills for text, images and layout, RAG ingestion including OCR, and connecting retrieval pipelines directly to workflows and agent tools. Multimodal extraction combining OCR, layout analysis and field extraction, clean grounded representations built with Content Understanding, and analyzers that produce structured or markdown output for downstream reasoning.

Written as a decision exam, not a service-API exam:

The audience profile assumes you develop in Python, and then the paper does not ask you to write any. What it asks, over and over, is which one and why: which class of model fits a described workload, which Foundry service does the grounding, which deployment option suits the load, which safety control catches the described failure, what a rate limit does to a batch job that shares a deployment with an interactive app. 404 of the 700 questions are written as a situation with a named team, engineer or organization in it rather than as a bare definition, and it is a stated constraint that decides the answer rather than a fact you can recall. That is the character of the exam, and it is also where preparation goes wrong. Anyone arriving from AI-102 revises service endpoints, resource keys and SDK method names, because that is what AI-102 rewarded. Very little of that is worth points here. What replaces it is judgment about a platform: what Foundry gives you, what it costs, what it logs, and what it stops.

Coming from AI-102:

AI-102 and the Azure AI Engineer Associate certification retired on 30 June 2026, and AI-103 leads to Azure AI Apps and Agents Developer Associate in its place. Treat it as a different exam rather than a refresh. The vocabulary carries over and so does most of your Azure knowledge, but the weight has moved. Agents were a 5 to 10 percent group on AI-102 and now anchor the largest domain on the paper, alongside orchestration, grounding and evaluation. Computer vision is image and video generation, editing and multimodal understanding, where AI-102 asked for image analysis, custom classification models, object detection, OCR and Video Indexer. Extraction is built on Content Understanding rather than Document Intelligence. And speech has not gone anywhere, it has simply ended up in a domain called text analysis, which is where people stop looking for it. One naming point is worth carrying into the exam room: the platform is Microsoft Foundry, and any material still calling it Azure AI Foundry predates December 2025.

Short scenarios with a stated constraint:

That is a deliberate match for how the objectives are written: they ask you to choose, configure and govern, so a question that made you parse a code listing would be testing something the exam does not. The work is in constraint. A budget, a latency target, a compliance boundary, a quota, a language requirement or an audit obligation appears in the stem, and it is what separates two options that both look defensible.

Exam Coach AI, besides every question:

Ask what agentic retrieval is, or when to ground a model instead of fine-tuning it, or what content filter blocks, while the clock is running. Exam Coach AI explains the concept, the distinction or the trap you just walked into, and it will not tell you which option to pick, which is the point. It is PrepifyLabs' own and it is tuned for AI-103 rather than being a general assistant bolted on, so it answers in the vocabulary this exam uses.

What AI-103 assumes you already know about Azure:

The audience expects familiarity with Azure services, and the exam takes that seriously. 196 questions, 28% of this bank, sit in a domain that is Azure administration applied to AI work: managed identity instead of keys, private networking, role policies, quotas and rate limits, cost footprint, and monitoring. If assigning a role, reading a metric or reasoning about a private endpoint is unfamiliar, that domain will cost you before you reach a single question about agents. Our AZ-104 Azure Administrator practice exam covers exactly that ground with 603 questions, and it is the sensible companion paper. If the design side is where you are weaker, AZ-305 Azure Solutions Architect covers architecture decisions across 700 questions.

How it is marked:

You need 70% to pass, scored as the plain percentage of questions you answered correctly out of 100%. After you submit, results break down by domain and by objective, so revision starts from a measured weak area rather than a guess. Each sitting draws 60 questions from the pool of 700, which is eleven distinct sittings before repetition starts to matter.

Free preview:

40 of the 700 questions are open without buying, spread across all five domains at 11, 13, 6, 5 and 5, in the same proportions as the exam itself, so the free sample has the shape of the paper rather than being its easiest corner.

Explore PrepifyLabs

Certification practice exams

Insights and case studies

The most advanced examination and certification readiness platform available. Train autonomously, pass effortlessly.

Prepifylabs LLC, 5900 Balcones Drive STE 38508, Austin, TX 78731, United States