AI-103 Practice Tests: Azure AI Apps & Agents Developer
250 exam-style questions with per-answer explanations and Microsoft Learn sources, current to the April 2026 skills
**Five full-length practice tests. 250 unique questions. Every answer explained, every question sourced.**
Exam AI-103, Developing AI Apps and Agents on Azure, replaced the retired AI-102 and leads to the Microsoft Certified: Azure AI Apps and Agents Developer Associate credential. It is a different exam, not a renumbering: Custom Vision, Video Indexer, CLU, custom question answering, prompt flow and knowledge store projections are all gone, while agents, tool integration, multi-agent orchestration, approval flows, observability and responsible AI carry far more weight than before.
These practice tests are written against the skills measured as of 16 April 2026, using current Microsoft Foundry naming rather than the older Azure AI Foundry and Cognitive Services terminology you will still find in most study material. Nothing here tests material that was dropped when AI-102 retired.
**What makes these different**
Every answer option has its own explanation, not just the correct one. When you pick a wrong answer, the feedback tells you what that option actually is, so the distinction sticks instead of the answer. Each question then closes with a full explanation of the reasoning and a link to the Microsoft Learn page it is based on, so you can read further where you are weak.
Difficulty rises deliberately across the five tests, from single-concept recall through service selection and genuine trade-offs to full multi-constraint exam simulation. The domain mix in every test matches the official weightings exactly, so your per-domain results add up across all five and tell you where to study.
**A note on scoring**
Microsoft reports a scaled score of 700 on a range of 1 to 1000 and states explicitly that this is not the same as 70% of the points, and that an easier question set requires more points. Many practice courses convert 700 into a 70% threshold, which is simply wrong. These tests set the bar above the real exam instead: 80% while you build up, 85% on the final simulation, because practising at your own desk is easier than sitting in a test centre.
**Exam facts covered in this practice set**
AI-103 runs 120 minutes with interactive components, multi-part questions and unscored trial questions, and there is no penalty for guessing. The certification is valid for one year and renewed through a free online open-book assessment.
Groundwork: Core Concepts and Foundry Fundamentals
50 questions · 80 minutes · pass mark 80%
Start here. Fifty foundation-level questions covering the core vocabulary of the AI-103 exam: deployment types and data residency, quota and throttling, keyless authentication and the Foundry roles, content filter categories and severities, the project endpoint and SDK client, agent tools, retrieval-augmented generation, image generation and vision input, text analysis and speech modes, and the building blocks of Azure AI Search and Document Intelligence. One concept per question with a short stem, so you can find out quickly which domains you already know and which need work. Every answer option has its own explanation and every question closes with a link to the Microsoft Learn page it came from. 80 minutes, 80% to pass.
In Practice: Choosing the Right Service
50 questions · 75 minutes · pass mark 80%
Fifty application-level questions built around short practical scenarios. You choose between services and settings that are easy to confuse: Data Zone versus Global, batch versus provisioned, managed compute versus serverless, blocklists versus Prompt Shields, hybrid versus vector retrieval, fast versus batch transcription, synchronous versus asynchronous document translation, template versus neural extraction models. This is the test that turns recognition into decision-making. Per-option feedback names what each wrong answer actually is, so a mistake teaches you the distinction rather than just the answer. 75 minutes, 80% to pass.
Trade-offs: When Two Answers Look Right
50 questions · 70 minutes · pass mark 80%
Fifty questions where more than one option is defensible and a single constraint decides: data residency, cost, latency, compliance boundary, whether citations are required, whether the wording of a source has to be preserved. This is where most AI-103 candidates lose marks, because the exam rarely asks which service exists and often asks which one fits. Covers deployment and spillover economics, network isolation, content filter approval paths, prompt engineering versus retrieval versus fine-tuning, structured outputs, reasoning models, agent tool selection under constraints, Content Understanding versus Document Intelligence, and multimodal ingestion strategies. 70 minutes, 80% to pass.
Under the Hood: Parameters, Limits and SDK Details
50 questions · 65 minutes · pass mark 80%
Fifty implementation-level questions on the details you only remember once you have written the code: parameter ranges and defaults, service limits, the order of operations, the property and header names that configure a feature. Covers spillover configuration, token scopes and private endpoint sub-resources, evaluator scales and thresholds, tool choice and approval policies, strict JSON schema constraints, image and video generation parameters, speech audio limits and formats, reciprocal rank fusion and reranker score ranges, vector index settings, and Document Intelligence add-on capabilities. Sit this one after you have worked through the labs. 65 minutes, 80% to pass.
Exam Day: Full-Length Simulation
50 questions · 60 minutes · pass mark 85%
A full-length simulation at real exam pressure: fifty questions in 60 minutes. That is deliberately tighter than the 120 minutes AI-103 allows, because the real exam spends much of its clock on labs, interactive components and multi-part questions that a text question does not. Matching the number would make this easier than the exam, not equal to it. Long multi-constraint scenarios that cross domains within a single question, the way the real exam does. Use this as a readiness gate rather than a study tool, which is why the bar is set at 85% - practising at home is milder than a test centre, so clearing it comfortably here is the signal you are ready to book. Note that Microsoft's own passing score is 700 on a scaled range of 1 to 1000, which is explicitly not the same as 70% of the points. 60 minutes, 85% to pass.
What you will practice
- Choose the right deployment type for a residency, latency or cost requirement, and explain what data at rest and inference routing actually mean
- Secure a Foundry solution with managed identity, keyless authentication, private networking and the correct Foundry role at the correct scope
- Build and operate agents: tool selection, function calling, MCP, approval flows, multi-agent delegation, and the limits each of them imposes
- Tell content filtering apart from evaluation, and configure both: harm categories and severities against evaluator scales and thresholds
- Apply retrieval correctly: hybrid search, semantic ranking, vector configuration, agentic retrieval and multimodal ingestion strategies
- Work confidently with image and video generation, inpainting, vision input and Content Understanding analyzers
- Select the right speech, language and translation capability for a scenario, and know the limits that decide between them
- Sit the real exam knowing your weak domains, because per-domain reporting is consistent across all five tests
Exam domains covered
Who is this for
- Developers and AI engineers preparing to sit exam AI-103
- Holders of the retired AI-102 certification moving to the current credential
- Azure practitioners who want to test what they actually know about Foundry, agents and grounding before booking an exam
- Anyone who wants explanations that teach the distinction rather than just marking an answer wrong
Requirements
- Experience developing applications, ideally in Python, as Microsoft expects for this exam
- Working familiarity with Azure and with generative AI concepts
- No prior certification is required, and AI-102 is not a prerequisite - it has been retired
Free exam guide
Not sure what to expect? Read how the exam is built, what it covers and how to prepare.