How to Pass the AI-103 Exam in 2026: Azure AI Apps and Agents Developer Guide

Last updated: June 2026. Researched against the official Microsoft Learn AI-103 study guide and certification pages. Examinotion fact-checks every exam article against Microsoft's published skills outline before publishing.

TL;DR Pass AI-103 by treating it as a Microsoft Foundry exam, not an Azure AI services exam. Generative AI and agentic solutions make up 30 to 35 per cent of the questions, so build agents and retrieval-augmented generation pipelines hands-on. The pass mark is 700 out of 1000, you get 120 minutes, and Python is the stated language.

Exam AI-103: Developing AI Apps and Agents on Azure is Microsoft's replacement for AI-102, and it earns the Microsoft Certified: Azure AI Apps and Agents Developer Associate certification [1]. It is the path forward for any developer building AI solutions on Azure, because AI-102 retires on 30 June 2026 [2]. This guide walks through what the exam tests, how it differs from AI-102, and a realistic plan to prepare without wasting effort on the wrong material.

AI-103 is a genuinely different exam, not a refreshed version of its predecessor. The platform, the tooling names, and the balance of topics have all moved on, and recycled AI-102 study notes will work against you. We will be honest about the difficulty throughout: this is one of the more demanding Azure AI associate exams, and preparation reduces your risk rather than guaranteeing a pass.

What is the AI-103 exam?

AI-103 validates your ability to design, build, and deploy AI applications and agents on Azure using Python and Microsoft Foundry [1]. The official audience profile describes the candidate as an Azure AI engineer who builds, manages, and deploys agents and AI solutions that take advantage of Microsoft Foundry [1]. It sits at the Associate (intermediate) level, the same tier as the AI-102 exam it replaces.

The exam is scenario-led rather than recall-led. Most questions present a situation and ask which Foundry service, agent pattern, or retrieval configuration is the right fit, so understanding why a service exists matters more than memorising its API surface.

Here are the verified fast facts as published by Microsoft.

Microsoft has not published an exact question count or a single fixed price for AI-103, so treat the figures above as the standard associate-exam pattern rather than AI-103-specific guarantees. The duration, passing score, and Python requirement are confirmed directly from the official pages.

Why AI-103 matters now: the AI-102 retirement

AI-102 retires on 30 June 2026, and after that date you cannot schedule or sit it [2]. If you already hold the AI-102 certification it stays on your Microsoft Learn transcript, but you will not be able to renew it once the exam is retired [2]. For anyone aiming at a current Azure AI associate credential, AI-103 is now the only forward route.

The timing creates a simple decision. If you are deep into AI-102 preparation and ready to sit it before the deadline, finishing it is reasonable because the certification remains valid. If you are starting fresh, there is little sense in preparing for an exam that disappears within days, so AI-103 is the better investment of your study time.

A Microsoft community expert put the difference bluntly in the official Q&A, describing AI-103 as "a very different exam for which you need to prepare for separately" [5]. Candidates coming from AI-102 should plan for real additional study rather than a quick revision top-up.

What changed from AI-102 to AI-103

Three shifts define the new exam: the move to a single platform, the rise of agentic AI, and retrieval-augmented generation becoming a core skill. Understanding these shifts tells you where the marks are.

Everything now runs through Microsoft Foundry

Microsoft Foundry is the unified platform that consolidates what used to be Azure AI Studio, the standalone Azure AI Services, and the Azure AI Foundry portal into one place [6]. In AI-103, Foundry is not one service among many. Every domain in the exam operates inside it, and resources are accessed through a Foundry project endpoint [6].

The rebranding matters because AI-102-era material uses names that no longer appear in the exam. Azure AI Studio and Azure AI Foundry are now Microsoft Foundry, the bundled Azure AI Services are now Foundry Tools, and the older Assistants API has been superseded by the Responses API for agents [6]. If your study notes mention Azure AI Studio as a standalone portal or the LUIS service by name, they predate AI-103.

Agentic AI moved from the edges to the centre

Building agents is now the single largest topic on the exam. The domain "Implement generative AI and agentic solutions" carries a 30 to 35 per cent weighting, larger than the entire generative AI coverage in AI-102 [1]. Multi-agent orchestration, agent safeguards and approval flows, autonomous and semi-autonomous workflows, and agent monitoring are all explicitly tested skills [1].

The Foundry Agent Service is the runtime you need to know, covering both prompt agents and hosted agents [7]. Expect questions that ask you to define agent roles, design tool schemas, integrate function-calling and conversation memory, and add oversight controls to keep an autonomous workflow safe [1].

RAG became a named, tested skill

Retrieval-augmented generation, or RAG, is the pattern of grounding a model's answers in your own data rather than relying on what the model already knew. In AI-103 it appears as an explicit skill in two separate domains: implementing RAG in an application, and configuring the RAG ingestion flow including documents and optical character recognition [1].

Azure AI Search underpins this, and its agentic retrieval capability, which runs a multi-query pipeline with semantic reranking, is now in scope [8]. A frequent exam trap is choosing fine-tuning when the scenario actually calls for RAG. Fine-tuning adapts a model's tone and behaviour; RAG injects private knowledge at inference time, and grounding questions almost always want RAG.

The five AI-103 skill domains and their weightings

The official skills outline divides AI-103 into five domains [1]. Your study time should follow the weightings, not split evenly across topics. The table below shows where the marks sit and where to concentrate.

Domain 1: Plan and manage an Azure AI solution (25 to 30%)

This domain is about architecture and operations. You choose the appropriate Foundry services and models for a task, design the Azure infrastructure, configure deployments, and integrate Foundry projects with CI/CD pipelines [1]. It also covers managing quotas, scaling, and cost, monitoring model performance, drift, and grounding quality, and securing systems with managed identity, private networking, and keyless credentials [1].

Responsible AI lives here too, tested at configuration depth rather than as principle. Expect questions on safety filters, guardrails, content moderation, trace logging, and governing agent behaviour with oversight modes and tool-access controls [1]. The featured Azure services include Microsoft Foundry, Microsoft Entra, Azure AI Search, Azure Monitor, and Azure AI Content Safety.

Domain 2: Implement generative AI and agentic solutions (30 to 35%)

This is the heart of the exam. You deploy and consume large, small, code, and multimodal models, implement RAG, and design tool-augmented flows and multistep reasoning pipelines using the Foundry SDKs [1]. You build agents that integrate retrieval, function-calling, and conversation memory, orchestrate multi-agent solutions, and add safeguards and approval flows to autonomous workflows [1].

Operationalising these systems is also tested: tuning generation behaviour through prompt engineering and model parameters, implementing self-critique loops, and setting up observability with tracing, token analytics, and latency breakdowns [1]. The key services are Microsoft Foundry, the Foundry Agent Service, Azure OpenAI, Azure AI Search, Prompt Flow, and Model Context Protocol servers for agent tools.

Domain 3: Implement computer vision solutions (10 to 15%)

This domain covers generative image and video workflows, multimodal understanding, and responsible AI for visual content [1]. You implement solutions that generate and edit images and video, build captioning and visual question-answering, and generate accessibility alt-text aligned to guidelines [1].

The notable new addition is Azure Content Understanding, a Foundry Tools service that did not exist in AI-102. You configure single-task and pro-mode Content Understanding pipelines to extract visual characteristics from images and video [1]. Domain 3 also expects you to detect and mitigate indirect prompt injection hidden in images, a modern threat that older material will not cover.

Domain 4: Implement text analysis solutions (10 to 15%)

Text analysis blends classic Azure AI Language capabilities with LLM-powered approaches, surfaced through Foundry Tools [1]. You extract entities, topics, summaries, and structured JSON using generative prompting, detect sentiment and sensitive content, and translate text with Azure Translator or LLM-powered flows [1].

A commonly missed corner is speech as an agent modality. The domain expects you to convert speech to text and text to speech for agentic interactions, integrate custom speech models, and enable multimodal reasoning from audio inputs [1]. Candidates who skip speech because it feels peripheral give away easy marks.

Domain 5: Implement information extraction solutions (10 to 15%)

This domain is the data plumbing behind RAG and agents. You ingest and index documents, images, audio, and video, configure semantic, hybrid, and vector search for grounding, and connect retrieval pipelines directly to agent tools [1]. Configuring the RAG ingestion flow, including OCR, is an explicit skill here [1].

Document extraction is the second half. You combine OCR, layout analysis, and field extraction in multimodal pipelines, and use Content Understanding to produce clean, grounded, markdown or structured outputs for downstream reasoning [1]. Azure AI Search and Azure AI Document Intelligence are the workhorses of this domain.

How to study for AI-103

The single most reliable principle is that hands-on Foundry time beats passive reading. AI-103 questions are architectural, and you cannot reason about which service fits a scenario if you have never built with any of them. Spin up a Foundry project, deploy a model, build a simple agent, and wire up a RAG pipeline before you touch a practice test.

The realistic timeline depends entirely on your starting point. The estimates below are Examinotion's preparation guidance based on the skills outline, not official Microsoft figures, and your own pace may differ.

Microsoft publishes free preparation paths and an exam sandbox to explore the question format, both linked from the official study guide [1]. Work through the Microsoft Learn training modules for each domain, then reinforce the high-weight Domain 2 material with your own Foundry builds. Examinotion's AI-103 practice tests and the AI-103 study guide are built to mirror the current skills outline so you drill the right topics rather than AI-102 leftovers.

If you are mapping out where this certification fits in your wider plan, our Microsoft AI certification roadmap shows how the fundamentals, associate, and architect exams stack together, and our guide to whether Microsoft AI certification is worth it covers the career context honestly.

Common reasons candidates fail AI-103

The pitfalls below are drawn from the structure of the official skills outline and the patterns seen across Azure AI associate exams. They are directional guidance, not a guarantee of what you will see.

Confusing Foundry components in scenarios. Candidates blur model deployment, the Foundry Agent Service, Prompt Flow, and Azure AI Search when several seem plausible. Each has a distinct role: Azure AI Search with agentic retrieval is the grounding layer, the Foundry Agent Service is the agent runtime, and Prompt Flow is for designed multistep reasoning with explicit flow control.

Picking the wrong extraction service. Azure AI Document Intelligence handles structured forms and tables through OCR and layout. Azure Content Understanding handles natural-language schema extraction across documents, images, audio, and video. Content Understanding is new in AI-103, and choosing Document Intelligence for multimodal sources is a frequent error.

Studying responsible AI at surface level. The exam tests specific configurations, which content filter category and severity blocks given content, how prompt shields work, what protected material detection covers, not high-level ethics. Learn the operational settings, not just the principles.

Choosing fine-tuning instead of RAG. When a scenario needs answers grounded in private or organisation-specific data, the answer is RAG, not fine-tuning. This distractor appears repeatedly.

Using outdated AI-102 material. Any resource that references Azure AI Studio as a standalone portal, the LUIS service, the Assistants API, or Azure Bot Service as a current topic is misaligned with AI-103. Vet every study note against the current skills outline.

Underweighting agents. The 30 to 35 per cent agentic domain dwarfs the 10 to 15 per cent vision, text, and extraction domains. Spreading study time evenly leaves you underprepared for the questions most likely to appear.

Exam-day logistics

AI-103 runs for 120 minutes, and the pass mark is 700 out of 1000 on Microsoft's scaled scoring [1][3]. The page also notes you may have interactive components to complete, so expect more than plain multiple choice [3].

Because AI-103 is an associate role-based exam, you can access Microsoft Learn during the exam [4]. This is a genuine advantage, but with limits: the timer keeps running while you read, access is restricted to the learn.microsoft.com domain, and the Q&A forums, practice assessments, and your personal profile are all blocked [4]. The practical lesson is not to lean on Learn access as a substitute for knowing the material. Practise navigating the Foundry documentation quickly so a lookup costs you seconds, not minutes.

Plan to renew. Microsoft associate certifications expire annually and are renewed by passing a free online assessment on Microsoft Learn before the expiry date [1]. There is no charge to renew, only the assessment.

Frequently asked questions

What is the AI-103 passing score?

The AI-103 passing score is 700 out of 1000 [1]. Microsoft uses scaled scoring rather than a flat percentage, so the threshold stays at 700 regardless of how difficult your particular question set is. Aim to score consistently above 80 per cent on practice assessments before you book the real exam.

How long is the AI-103 exam?

AI-103 gives you 120 minutes to complete the assessment [3]. That is 20 minutes longer than the retired AI-102, reflecting the more complex, scenario-based questions on agentic and generative AI architectures. Budget your time so the high-weight Domain 2 questions get the attention they deserve.

Does AI-103 require Python?

Yes. Microsoft's official audience profile states candidates should have experience developing apps using Python [1]. C# is supported in the Azure AI SDKs and may appear as context, but Python is the stated requirement. If you have only worked in C#, practise the Python Foundry SDK calls before sitting the exam.

Can I still take AI-102 instead?

AI-102 retires on 30 June 2026, after which you cannot schedule or sit it [2]. Until then you can still take it, and any AI-102 certification you already hold remains on your transcript, though it cannot be renewed after retirement [2]. For a current credential, AI-103 is the path going forward.

Is AI-103 hard?

AI-103 is one of the more demanding Azure AI associate exams, but its difficulty is architectural rather than memory-based. It asks which Foundry service, agent pattern, or RAG configuration suits a scenario, so hands-on experience helps far more than rote study. The 30 to 35 per cent agentic domain is where underprepared candidates most often lose marks.

What is Microsoft Foundry and why does it matter for AI-103?

Microsoft Foundry is the unified Azure platform for enterprise AI development, consolidating the former Azure AI Studio, Azure AI Services, and Azure AI Foundry portal into one [6]. In AI-103 every domain operates within Foundry, so understanding its structure of resources, projects, the model catalogue, the Agent Service, and Foundry Tools is foundational to passing.

Conclusion

AI-103 rewards builders. Treat it as a Microsoft Foundry exam, weight your study toward agents and RAG, and verify every study note against the current skills outline rather than trusting AI-102-era material. The pass mark is 700 out of 1000 in 120 minutes, Python is the working language, and you can lean on Microsoft Learn during the exam as long as you already know your way around.

When you are ready to test that knowledge under exam conditions, start practising for AI-103 with Examinotion's practice tests, or browse the full Microsoft AI exam catalogue to plan your wider certification path. If you are weighing your options, our AI-901 exam guide covers the fundamentals stepping stone, and our Azure AI services comparison goes deeper on the underlying platform.

Sources

1. Exam AI-103 study guide — Microsoft Learn, accessed 2026-06-22
2. Retired certification exams — Microsoft Learn, accessed 2026-06-22
3. Azure AI Apps and Agents Developer Associate certification — Microsoft Learn, accessed 2026-06-22
4. Exam duration and exam experience — Microsoft Learn, accessed 2026-06-22
5. AI-102 retires on 30th June 2026 (community thread) — Microsoft Q&A, accessed 2026-06-22
6. What is Microsoft Foundry — Microsoft Learn, accessed 2026-06-22
7. Foundry Agent Service overview — Microsoft Learn, accessed 2026-06-22
8. Agentic retrieval in Azure AI Search — Microsoft Learn, accessed 2026-06-22