Exam AI-103, Developing AI Apps and Agents on Azure, earns the Microsoft Certified: Azure AI Apps and Agents Developer Associate credential. Microsoft writes it for developers who build on Microsoft Foundry: they pick models and deployments, wire retrieval and tools into agents, put guardrails and evaluations around what they ship, and add vision, speech, text analysis and document extraction where an app needs them. You are expected to have built apps in Python, and most questions read like a design review, with a requirement on the table and four ways to meet it.
The timing matters. AI-102, the Azure AI Engineer Associate exam, retired on June 30, 2026, and a large share of the study material online was written for it. Services have been renamed since then, the agent API has moved, and some features that AI-102 guides teach have been retired. This guide covers what AI-103 tests today, who it suits, how it sits next to its neighbors, and a plan for preparing in about four weeks.
Exam Quick Facts
Find out where you stand first
Take the free AI-103 practice questions cold, with no signup, and read on with your gaps in mind. For full rehearsal, preporato.com has six timed AI-103 practice exams with an explanation for every option, and an AI-103 course that walks the five skill areas in exam order.
How AI-103 questions are built
AI-103 rewards precise product knowledge applied to a stated requirement. A typical item describes an app, an agent or a pipeline, names one constraint, and asks which service, setting or design meets it. The constraint might be that data stays in a geography, that an identity holds the least privilege it needs, that a tool never runs without a person's approval, or that an answer can be traced to its source. Several options usually work in a general sense, and only one of them meets the constraint the question states.
Three habits separate passing scores from near misses. Read for the constraint before you read the options. Know the exact control that enforces a requirement: an instruction in a system prompt asks an agent to behave, while allowed_tools, require_approval and a deployment guardrail make it behave. And know what each feature needs before it works, such as which evaluators need a reference answer, which role an identity needs on which resource, and which retrieval options exist for agents.
Microsoft states that most questions cover generally available features and that preview features can appear when they are commonly used. A feature marked preview in the documentation is still fair game.
Preparing for AI-103? Practice with 390+ exam questions
The five skill areas
The study guide, with skills measured as of April 16, 2026, splits the exam into five areas. The first two carry well over half of the exam between them. The last three are smaller and full of specific services that are easy to confuse under time pressure. For every objective matched to the Foundry services and settings that answer it, see AI-103 exam topics explained.
Core Topics
- •Choosing models: large and small language models, multimodal models and Foundry Tools
- •Choosing Foundry services for generation, grounding, vector search and agent workflows
- •Retrieval and indexing methods, plus memory, tool and knowledge services for agents
- •Infrastructure, deployment options and Foundry projects in CI/CD pipelines
- •Quotas, scaling, rate limits and cost for model and agent workloads
- •Monitoring drift, safety events, grounding quality and search index health
- •Managed identity, private networking, keyless credentials and role policies
- •Guardrails, evaluators, trace logging, approval workflows and agent oversight
Skills Tested
Example Question Topics
- Traffic spikes past provisioned capacity at month end. What keeps requests flowing?
- Which role lets an app call deployments without managing the resource?
- A tool can issue refunds. How does a person approve each call?
What moved since AI-102
Most AI-103 mistakes trace back to study material that describes the platform as it was. These are the changes that show up in answer options.
AI-102 era material and the platform AI-103 tests
| Area | Older material describes | AI-103 expects |
|---|---|---|
| Platform | Azure AI Studio and Azure AI Foundry hubs | Microsoft Foundry: a Foundry resource with projects |
| Prebuilt services | Azure AI services under their old names | Foundry Tools, such as Azure Speech in Foundry Tools |
| Agents | Assistants API threads, messages and runs | Agents on the Responses API, with versions, conversations and tools such as MCP |
| API versions | A monthly api-version parameter | Stable v1 routes under /openai/v1/ |
| Image generation | DALL-E 3 | GPT-image models for images and Sora 2 for video |
| Reasoning over documents | Content Understanding pro mode | Agentic mode, in preview on the 2026-06-01-preview API |
| Retrieval for agents | A search index called from code | Agentic retrieval and Foundry IQ knowledge bases |
One line in the study guide still names single-task and pro-mode Content Understanding pipelines. Pro mode retired with the 2025-05-01-preview API, and Microsoft's documentation now sends that reasoning-heavy work to agentic mode, which handles one input file per request and has no extract fields. Learn both names and what each one does, so the older term in a question doesn't throw you.
Older notes still help with concepts that changed least, such as how Azure AI Search indexes and ranks content. Check every service name and API against current Microsoft Learn pages before you memorize it. The AI-103 vs AI-102 comparison walks through the changes job by job, with what to relearn first.
Who AI-103 is for
Microsoft describes the candidate as an Azure AI engineer who builds, manages and deploys agents and AI solutions on Microsoft Foundry, working with business stakeholders, solution architects, data scientists, DevOps engineers and cloud security engineers. In practice that is a developer who already ships on Azure and now owns the generative side of an app: the model choice, the agent, the retrieval behind it and the controls around it.
It is a good fit if you write Python, have used the Foundry portal or SDK at least a little, and want a credential that matches agent and RAG work. It is a harder start if you have never deployed anything to Azure, since many questions assume you know how identities, networks and resources fit together.
AI-103 also has neighbors that are easy to mix up when choosing what to take.
AI-103 and related Microsoft certifications
| Certification | Built for | Choose it when |
|---|---|---|
| AI-103: Azure AI Apps and Agents Developer Associate | Developers who build AI apps and agents on Microsoft Foundry | You design and ship the generative and agent layer of an app |
| AI-200: Azure AI Cloud Developer Associate | Developers who build the back-end services of AI solutions: Azure SDKs, data services, messaging, vector databases and containers | Your work is the services and data plumbing around AI features |
| AI-300: Machine Learning Operations Engineer Associate | Engineers who run MLOps and GenAIOps with Azure Machine Learning, Foundry, GitHub Actions and Bicep | You own training, deployment pipelines and monitoring for models and AI apps |
| Agentic AI Business Solutions Architect Expert (exam AB-100) | Architects who design AI-powered solutions across Microsoft AI and business applications | You already hold a qualifying associate certification, and AI-103 is one of them |
Three AI-103 questions
A support agent in Foundry Agent Service connects to an MCP server that exposes 40 tools. The agent should reach only three of them: order lookup, shipping status and refunds. A person must confirm every refund before it runs. Which configuration meets both requirements?
Master These Concepts with Practice
Our AI-103 practice bundle includes:
- 6 full practice exams (390+ questions)
- Detailed explanations for every answer
- Domain-by-domain performance tracking
30-day money-back guarantee
A four-week plan
For developers who already build on Azure, plan on 40 to 60 hours over four to six weeks. The schedule below fits four weeks at 10 to 15 hours a week, following the order of the AI-103 course. For a hands-on build to pair with each skill area and a plan for exam day, see how to pass AI-103 on your first attempt.
Week 1: baseline, then plan and manage. Take one practice exam cold before studying anything, and keep the per-area scores. Then work through the Foundry resource and project model, deployment types, quotas and spillover, identity and private networking, and responsible AI controls: guardrails, evaluations and agent oversight.
Week 2: generative apps and agents. Connect an app to a Foundry project with the SDK, keep conversation state with the Responses API, build RAG, and give an agent tools through functions, OpenAPI and MCP, with memory and structured inputs.
Week 3: orchestration, evaluation and retrieval. Multi-agent patterns in Microsoft Agent Framework, human-in-the-loop controls, quality and agent evaluators, then Azure AI Search indexing, hybrid and semantic ranking, vector design, agentic retrieval and Content Understanding.
Week 4: vision, language, speech and rehearsal. Image and video generation, visual analysis, text analysis, translation and speech, then the remaining practice exams under time, reviewing every explanation, including the ones for questions you answered correctly.
Pacing
Score every practice exam by skill area and spend the next study block on the weakest area. When the same kind of question keeps costing marks, such as roles or evaluator inputs, write a one-line rule for it and reread your rules before the next exam. Microsoft's free Practice Assessment on AI Skills Navigator and the exam sandbox are worth one pass each in the final week.
Mistakes that cost marks
Studying from AI-102 material. Hub-based projects, the Assistants API, DALL-E 3 and pro mode all appear in older guides. Options built on them can sound right and describe a platform that has moved on.
Answering with an instruction where the question wants a control. If a scenario says an agent must never call a tool without approval, or must never leave its task, look for allowed_tools, require_approval, a guardrail control or an approval workflow. A sentence in the system prompt is a request the model can ignore.
Choosing the most capable option over the one that fits. Agentic mode in Content Understanding earns its cost when a value has to be calculated or checked across a contract and its amendments. Reading an invoice number is a job for a standard analyzer, which is faster and cheaper. The semantic ranker reorders results that retrieval already found, so it cannot rescue a document that the query never matched.
Skipping identity and networking details. Keyless access with managed identities, the role an identity needs on which resource, private endpoints and the DNS zones behind them all carry marks. These questions are short and precise, and a vague answer loses them.
Frequently asked questions
Where to start
Read Microsoft's study guide first, since its skills list is the scope of the exam. Then take the free AI-103 practice questions without preparing, so the result tells you where to begin. For ten more exam-style scenarios with every option explained, work through the AI-103 practice questions.
The AI-103 course on preporato.com follows the five skill areas in exam order, with an exam tip and three checkpoint questions in every lesson. The six timed AI-103 practice exams rehearse the real thing, with an explanation for every option and a score for each skill area.
Sources:
- Microsoft Certified: Azure AI Apps and Agents Developer Associate
- Study guide for Exam AI-103: Developing AI Apps and Agents on Azure
- Exam and assessment lab retirement
- Renew your Microsoft Certification
- What is Microsoft Foundry?
- Agentic mode in Content Understanding
- Microsoft Certified: Azure AI Cloud Developer Associate (AI-200)
- Microsoft Certified: Machine Learning Operations Engineer Associate (AI-300)
- Microsoft Certified: Agentic AI Business Solutions Architect Expert
Ready to Pass the AI-103 Exam?
Join thousands who passed with Preporato practice tests