Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI‑103) Practice Exams
About the Azure AI-103 exam
Exam at a glance
AI-103 is Microsoft's associate certification for developers who build generative AI and agentic applications on Azure. Passing it earns the Microsoft Certified: Azure AI Apps and Agents Developer Associate credential. The exam is hands-on and Python-first: you are expected to implement solutions on Microsoft Foundry — Foundry Agent Service, Foundry Models, Azure AI Search, Content Understanding, Content Safety, Speech, and Translator — rather than simply describe them.
Domain weighting
- Plan and manage an Azure AI solution: 25–30%
- Implement generative AI and agentic solutions: 30–35%
- Implement computer vision solutions: 10–15%
- Implement text analysis solutions: 10–15%
- Implement information extraction solutions: 10–15%
Who this exam is for
AI-103 targets Azure AI engineers and application developers who design, build, deploy, and maintain AI apps and agents on Microsoft Foundry. You work in Python and collaborate with solution architects, data scientists, DevOps engineers, and security engineers. Questions are scenario-based: a short stem names a concrete Foundry workload and its constraints, and you pick the service, mode, deployment option, retrieval approach, or role that satisfies all of them.
Prerequisites
There are no formal prerequisites. Microsoft recommends experience developing apps with Python and familiarity with general AI, generative AI, and Azure AI services. You should be comfortable deploying models and agents in a Foundry project, grounding responses with Azure AI Search, and securing resources with managed identity and role-based access control. If you already hold AI-102 or AI-900, much of the service vocabulary will be familiar and your focus shifts to the agentic and Foundry-specific material.
Why take this certification
- The app-and-agent developer credential. AI-103 proves you can ship production generative AI and agents on Microsoft Foundry with Python, not just recognize the services on a slide.
- Foundry-native and Python-first. The blueprint is built around Foundry Agent Service, Foundry Models, Azure AI Search, Content Understanding, Content Safety, Speech, and Translator, exercised through the Foundry SDKs.
- Where AI hiring is moving. Agentic apps, retrieval-augmented generation, and multimodal understanding are exactly the skills teams are hiring for in 2026, and this exam is organized around them.
- Builder altitude, not trivia. Questions reward implementation judgment — choosing the right analyzer mode, retrieval strategy, deployment option, or least-privilege role — over memorized definitions.
What you'll learn in the AI-103 exam
AI-103 is hands-on and scenario-driven. Most questions describe a short Foundry implementation task where two or three services or modes all look plausible, and the better answer follows a single instinct that runs through the whole exam: prefer the managed, least-privilege, purpose-built capability that satisfies every stated constraint, and reject each near-miss on the one constraint it violates. Often the first move is naming the service and the exact mode or feature the scenario needs, and the configuration follows.
Core services and tools you'll be tested on
- Plan and manage an Azure AI solution: choosing models and Foundry services for generation, grounding, vector search, agent workflows, and multimodal processing; retrieval and indexing methods; memory, tool, and knowledge integration for agents; deployment options such as serverless API endpoints, managed compute, and Foundry Local; CI/CD for Foundry projects; quotas, scaling, rate limits, and cost; and security with managed identity, private networking, keyless credentials, and role-based access control.
- Implement generative AI and agentic solutions: deploying and consuming large, small, code, and multimodal models; retrieval-augmented generation; building agents with Foundry Agent Service using function calling, tools, MCP, and conversation memory; orchestrated multi-agent workflows with approval and safeguard controls; and evaluation and observability through tracing, token analytics, safety signals, and latency breakdowns.
- Implement computer vision solutions: generating and editing images and video from text prompts and reference media, including inpainting, mask edits, and remix versus input_reference controls; multimodal understanding and single- or multi-image captioning; visual question answering; Content Understanding analyzers for visual characteristics; and responsible AI for visual content, including Prompt Shields for injected text in images.
- Implement text analysis solutions: extracting entities, topics, summaries, and structured JSON with generative prompting and Foundry Tools; detecting sentiment, tone, and sensitive content; translating with Azure Translator or LLM-powered flows; and speech workflows for speech-to-text, text-to-speech, custom speech models, and speech translation.
- Implement information extraction solutions: ingesting and indexing documents, images, audio, and video; configuring semantic, hybrid, and vector search for grounding; enrichment with built-in and custom skills and OCR; connecting retrieval pipelines to agent tools; and Content Understanding document analyzers that produce structured or markdown output for downstream reasoning.
Implementation judgment patterns you'll need to recognize
- Choosing the managed, least-privilege, keyless option, such as a managed identity over a stored key, or the right Foundry role over broad Contributor rights.
- Discriminating between modes and options that all look valid: Content Understanding standard versus pro mode, agentic versus classic versus iterative retrieval, serverless API versus managed compute versus self-hosting, and real-time versus batch transcription.
- Completing the function-calling loop correctly by executing the tool and returning a function_call_output keyed to the matching call_id, rather than displaying the raw arguments to the user.
- Reaching for the right evaluator: AI-assisted quality evaluators need a judge-model deployment, while agent-specific risks such as Prohibited Actions and Indirect Attack are not caught by the four content-harm categories.
- Sizing and operating deployments, such as factoring the prompt-cache rate into provisioned throughput sizing and reading the spillover metric before resizing a deployment.
- Grounding reliably by choosing structured outputs over plain JSON mode, and the recommended Content Understanding skill over the legacy Document Layout skill for tables that span pages.
How the practice exams help
Each free question and every premium exam mirrors the scenario style Microsoft uses: a short stem with two or three constraints and several plausible Foundry options. Detailed explanations cover not just why the right answer is right but why each distractor fails on exactly one stated constraint, so you learn the trade-offs rather than memorizing answers.
How to prepare for the AI-103 exam
A successful AI-103 plan pairs structured study with daily hands-on time building on Microsoft Foundry in Python. A recommended approach:
- Study the skills measured (2 to 3 weeks). Walk the official AI-103 study guide and the Microsoft Learn training modules it links for each domain. Prioritize Implement generative AI and agentic solutions (30–35%) and Plan and manage an Azure AI solution (25–30%), which together account for more than half of the exam.
- Hands-on labs (3 to 4 weeks). Stand up a Foundry project and build in Python: deploy a model, wire retrieval-augmented generation over Azure AI Search, and create an agent with Foundry Agent Service that calls function tools and keeps conversation memory. Add Content Understanding document analyzers, Content Safety filters and Prompt Shields, and Speech and Translator flows. Practice deployment options, managed identity, and role-based access control as you go.
- Evaluation and observability (1 week). Run AI-assisted quality and risk-and-safety evaluators against a fixed dataset, wire tracing and token analytics, and set up continuous and scheduled evaluation, since measuring and governing agents runs through every domain and often decides between two otherwise valid answers.
- Practice exams (1 to 2 weeks). Take timed practice tests to surface weak areas. Detailed explanations on every option help you learn the reasoning rather than memorize answers. Aim for consistent scores above 80% before you schedule.
Recommended timeline
Plan on 6 to 10 weeks of focused study (8 to 12 hours per week) for developers who already have some Python and Azure experience. If you come from AI-102 or AI-900, you can move faster through the service fundamentals and should spend the extra time on Foundry Agent Service, agent orchestration, evaluation and observability, and the newer multimodal and Content Understanding features.
Official resources
Read the official AI-103 study guide for the exact skills measured, which were last updated on April 16, 2026, and review the AI-103 exam details page for registration and how to earn the Azure AI Apps and Agents Developer Associate certification. A Microsoft practice assessment is not yet published for this exam; it usually appears within about eight weeks of an exam leaving beta, so lean on hands-on labs and these practice tests in the meantime.