AI-103 study guide: domains, format and a 4-week plan

AI-103, Developing AI Apps and Agents on Azure, is the exam for the Microsoft Certified: Azure AI Apps and Agents Developer Associate certification. It is written for Azure AI engineers who build, deploy and manage generative AI apps and agents with Microsoft Foundry, and who develop in Python. You have 120 minutes, the exam is proctored and may include interactive components, and the passing score is 700.

Microsoft lists AI-103 as the replacement for AI-102 and the Azure AI Engineer Associate certification, which retired on 30 June 2026. If you started preparing for AI-102, expect a much heavier emphasis on agents, orchestration and operating AI systems in production.

AI-103 at a glance

Details as of September 2026 — confirm on the official exam page before booking.

Who this exam is for

Microsoft's audience profile describes an Azure AI engineer who builds, manages and deploys agents and AI solutions on Foundry, working with stakeholders, architects, data scientists, DevOps and security engineers. You are expected to have experience developing apps in Python and to be familiar with general AI, generative AI and Azure services.

There is no formal prerequisite exam. AI-901 is a sensible warm-up if Foundry is new to you, but the gap between the two is large: AI-103 assumes you have deployed models, written code against them and dealt with identity, networking and monitoring.

What the exam covers

The AI-103 study guide lists skills measured as of 16 April 2026 in five domains.

Plan and manage an Azure AI solution (25–30%)

This domain is about decisions and operations. You choose models (large, small, multimodal) and Foundry services for generation, grounding, vector search and agent workflows, plus the retrieval method and the memory, tool and knowledge services an agent needs. It then covers infrastructure and deployment options, CI/CD for Foundry projects, quotas and cost, monitoring for drift, safety and grounding quality, and security through managed identity, private networking, keyless credentials and roles.

The last task is responsible AI in practice: safety filters and guardrails, evaluators and safety evaluations, auditing with traces and approval workflows, and governing what agents are allowed to do. Treat these as configuration questions, not ethics questions.

Implement generative AI and agentic solutions (30–35%)

The heaviest domain. For applications, it covers deploying and consuming different model types, implementing RAG, designing multistep and tool-augmented workflows, evaluating output for fabrication, relevance and safety, and connecting an app to a Foundry project through the SDKs.

For agents, you define roles, goals, conversation tracking and tool schemas, integrate retrieval, function calling and memory, orchestrate multiple agents, and build semiautonomous workflows with approval controls. The optimisation task adds prompt and parameter tuning, self-critique loops, tracing and token analytics, and combining models with rules engines.

Implement computer vision solutions (10–15%)

This domain is now mostly generative and multimodal. It covers image and video generation from prompts and reference media, inpainting and mask-based editing, captions, visual question answering and accessible alt text, and Content Understanding for visual characteristics and video segments. It also has a responsible AI task: filtering unsafe images, detecting prompt injection hidden in text within images, and enforcing visual policies such as watermarks and brand rules.

Implement text analysis solutions (10–15%)

Text analysis on AI-103 is done largely with language models and Foundry Tools: extracting entities, topics, summaries and structured JSON, detecting sentiment and sensitive content, translating with Azure Translator or a language-model flow, and adapting output to a domain. The speech task covers speech-to-text and text-to-speech in agent interactions, custom speech models, reasoning over audio, and speech translation.

Implement information extraction solutions (10–15%)

Two tasks. The first is retrieval and grounding pipelines: ingesting and indexing documents, images, audio and video; semantic, hybrid and vector search; enrichment skills; OCR in the RAG ingestion flow; and connecting retrieval to agent tools. The second is document extraction with Content Understanding, combining OCR, layout and field extraction and producing structured or Markdown output for downstream reasoning.

A 4-week study plan

Roughly 60% of the marks sit in the first two domains, so they get two of the four weeks.

Week 1: plan and manage. Build a Foundry project from scratch with managed identity and keyless access, deploy two models of different sizes, and wire up monitoring. Configure guardrails and run a built-in evaluation. Write down why you chose each deployment option.

Week 2: apps and agents. Build a RAG app against your own documents, then an agent with a function tool, retrieval and conversation memory. Add a second agent and an approval step before a sensitive tool call. Turn on tracing and read a trace end to end.

Week 3: vision, text and speech. Split the week across the three 10–15% areas that involve media: generate and edit an image, caption and question an image with a multimodal model, extract structured JSON from text, translate a document, and run a speech-to-text and text-to-speech round trip.

Week 4: extraction and mocks. Build an index with hybrid search and an OCR step, and run Content Understanding on a scanned form. Then take a full-length mock, review every explanation, repair the weakest domain, and take a second mock.

Where candidates lose marks

• Studying AI-102 material. Older courses cover services and patterns that AI-103 no longer centres on. Check every resource against the April 2026 skills list.
• Using outdated Foundry terminology. Microsoft's Foundry documentation has changed quickly, and some pages still describe older agent concepts. Learn from current docs and reason from capabilities rather than memorised API names.
• Confusing search modes. Semantic, hybrid and vector search solve different problems. Scenario questions often hinge on whether exact keyword matches, meaning, or both matter.
• Treating guardrails as a checkbox. Questions ask what a filter, evaluator or approval workflow actually does in a given situation, and when a control applies at input, output or tool-call time.
• Relying on Microsoft Learn during the exam. Role-based exams give access to Learn, but no extra time is added. It helps with one lookup, not a dozen.
• Case study pacing. Case studies carry long scenarios. Read the requirements first, then the background, and do not start a break mid-set expecting to return.

Practising for AI-103 on CertifyAIx

CertifyAIx has 579 original AI-103 questions, organised by the exam's five domains and each explained option by option. Our [methodology](/blog/how-certifyaix-practice-questions-are-written) describes how they are written.

• 3 full-length mock exams, timed to the real exam's 120 minutes
• 26 mixed domain papers and 29 topic papers of up to 25 questions each (60 minutes per 25)
• Case-study, drag-and-drop, hot area and ordering formats alongside single and multiple choice

Start with the [AI-103 practice questions](/certifications/azure-ai-app-agent-developer-ai-103), and see the [AI certification study plan template](/blog/ai-certification-study-plan-template) if you want to adapt the plan above.

Frequently asked questions

How much experience do I need for AI-103?

Microsoft asks for Python development experience and familiarity with generative AI and Azure services. In practice, you should have built and deployed at least a few Foundry apps or agents before you sit it.

How long does AI-103 preparation take?

Four to eight weeks is a realistic range. Developers already building on Foundry can finish at the short end; those coming from AI-102 or AI-901 need longer, mainly for agents and operations.

Is AI-102 still available?

No. AI-102 and the Azure AI Engineer Associate certification retired on 30 June 2026, and Microsoft names AI-103 as the replacement. Microsoft says certifications already earned remain valid, and on your transcript, until they expire.

What happens if I fail?

Microsoft allows a retake 24 hours after a first failed attempt; waiting periods for later retakes are longer and are set out in its retake policy.

Are CertifyAIx questions taken from the real exam?

No. Every question is original, written against Microsoft's published study guide. We never use live or past exam content.

CertifyAIx is independent and not affiliated with Microsoft. Microsoft Certified: Azure AI Apps and Agents Developer Associate is a trademark of its owner. All CertifyAIx practice questions are original.