Home Certifications FAQ Partners Blog

·EN·DE·FR·ES·PT

Microsoft · AI-103 · Associate

Developing AI Apps and Agents on Azure (AI-103) — Practice Questions and Mock Exam

Prepare for AI-103 with original practice questions and clear answer explanations. Ask Alex, your AI tutor, when you need more detail, use your results to identify topics to review, and practice your pacing with timed mock exams.

Start practicing AI-103 free Buy AI-103 access for $29.99

Buy AI-103 access for $29.99 Start practicing AI-103 free

50Mock exam questions

120minTime limit

700/ 1000Passing score

Checked against Microsoft · August 2026 · Current exam version

Overview

About the exam

Exam AI-103, Developing AI Apps and Agents on Azure, leads to the Microsoft Certified: Azure AI Apps and Agents Developer Associate certification. It is the successor to AI-102, which is now retired, and shifts the focus from wiring up pre-built Azure AI services to building generative AI applications and multi-agent solutions on Microsoft Foundry.

The exam targets Azure AI engineers who design, build, deploy, and operate AI apps and agents using Foundry, Azure OpenAI models, Azure AI Search, Azure AI Content Understanding, and the Azure AI Language, Vision, and Speech capabilities. You are expected to develop in Python and to understand generative AI patterns such as retrieval-augmented generation (RAG), tool calling, agent orchestration, and responsible AI.

Try the practice preview

Try five AI-103 questions

Try five practice questions from the app’s current Developing AI Apps and Agents on Azure question bank, with answers and explanations.

Implement information extraction solutions1 / 5

You must ingest mixed PDFs that contain both narrative text and embedded charts/diagrams into an Azure AI Search index for multimodal RAG. You want a single skill that analyzes layout, produces Markdown that preserves section structure, and extracts images with location metadata so chunks stay near their related figures. Which built-in skill should you use?

AlexFull explanation from Alex

The Document Layout skill in Azure AI Search calls the Azure Document Intelligence layout model to detect document structure and emit a syntactical representation in Markdown or text. Its markdownHeaderDepth setting controls section nesting, and when configured for text output with image extraction it returns normalized_images plus locationMetadata that preserves each image's position in the document. Keeping images near their related content is valuable for RAG and multimodal search. The OCR skill, by contrast, only extracts text from images and does not produce structured Markdown sections; the Image Analysis skill describes image content but does not chunk a document by layout. Exam tip: 'structure-aware Markdown plus image location metadata for multimodal RAG' points to the Document Layout skill, not plain OCR.

Sourcelearn.microsoft.com

Implement generative AI and agentic solutions2 / 5

A platform team wants to debug a multi-step agent by seeing the full execution flow, including each LLM call, tool invocation, and agent decision, with data flowing into Application Insights. Which observability capability in Microsoft Foundry provides this end-to-end execution visibility?

AlexFull explanation from Alex

Microsoft Foundry's observability rests on three pillars: evaluation (measuring quality and safety), monitoring (production dashboards and alerts), and tracing. Tracing provides distributed, end-to-end visibility into an AI application's execution: it records LLM calls, tool invocations, agent decisions, and inter-service dependencies. It is built on the OpenTelemetry standard and integrates with Azure Monitor Application Insights, which is exactly what you need to debug multi-step agent behavior, find bottlenecks, and understand reasoning chains. Tracing supports frameworks including the Microsoft Agent Framework and the OpenAI Agents SDK. Fine-tuning, prompt flow variants, and content filtering address customization, prompt comparison, and safety respectively, not runtime execution visibility. Exam tip: when a question mentions 'execution flow', 'every decision', 'OpenTelemetry', or 'Application Insights for debugging agents', the answer is tracing.

Sourcelearn.microsoft.com

Implement text analysis solutions3 / 5

An application sends short user-generated comments to the Azure AI Translator v3 /translate REST endpoint. The team wants each request to translate into both German and Italian in a single call, and to let the service detect the source language automatically. Which approach correctly configures the request?

AlexFull explanation from Alex

The Azure AI Translator v3 /translate endpoint takes the target language(s) as the required 'to' query parameter. To translate into several languages in one call you repeat the parameter, for example to=de&to=it. The 'from' parameter is optional: when omitted, the service automatically detects the source language and returns the detected language and a confidence score in the response. You only need the separate /detect endpoint when you want detection without translation. Note that the dynamic dictionary feature requires you to specify 'from' explicitly rather than relying on autodetection. Exam tip: multiple targets are achieved by repeating 'to', and leaving out 'from' is the supported way to get inline auto-detection during translation.

Sourcelearn.microsoft.com

Plan and manage an Azure AI solution4 / 5

Your security policy requires that approval is obtained before any change, but a developer asks to set a deployment's content filter to 'No filters' (turn filtering off) for an Azure OpenAI model. What is required before this configuration can take effect?

AlexFull explanation from Alex

Foundry content filters are configurable: any customer can set the severity threshold (low, medium, or high) separately for prompts and completions. However, the 'No filters' and 'Annotate only' options, which disable blocking, require special approval. For Azure OpenAI models, only customers approved through the Azure OpenAI Limited Access Review for modified content filters can fully turn off filtering. This is a governance guardrail to keep responsible-AI protections in place by default. Exam tip: changing thresholds = no approval needed; disabling filtering entirely or annotate-only = requires the modified-content-filter approval.

Sourcelearn.microsoft.com

Implement computer vision solutions5 / 5

A solution must extract custom structured fields directly from product-shelf photos (for example, shelf-share percentages and detected brands) using a schema you define, including classifying images into categories and generating fields with a generative model. Which Azure capability is purpose-built for this image field extraction?

AlexFull explanation from Alex

Azure AI Content Understanding extends field extraction across all modalities, including images. Its image field extraction capability lets you extract custom structured fields from images based on a schema you define; fields can be extracted directly, classified from a set of categories, or generated using a generative model. This is well suited to retail shelf analysis, manufacturing quality control, and chart-based BI where you need structured outputs rather than free-form captions. Unlike Image Analysis Caption (a fixed sentence) or generic VQA, image field extraction is schema-driven and returns machine-readable fields. Exam tip: 'custom schema-driven structured fields from images, with classification and generative options' = Content Understanding image field extraction.

Sourcelearn.microsoft.com

Create a free account

Question pool

460 practice questions

The Pass-IT question pool gives you material to practice for AI-103. A Pass-IT mock exam uses 50 questions and a 120-minute time limit; these are practice settings.

Pool details: AI-103

Exam details checked against MicrosoftAugust 28, 2026

date of the last check against the official Microsoft source

Passing score700 / 1,000

as published by Microsoft

Objectives in the guide78 objectives listed in the official guide

across 5 domains in the official exam guide

Pool size460 questions

= The pool size is equivalent to 9 sets of 50 questions; this does not mean that each mock exam uses a separate set.

Blueprint domains5 domains in the exam blueprint

Plan and manage an Azure AI solution 123 · Implement generative AI and agentic solutions 153 · Implement computer vision solutions 59 · Implement text analysis solutions 60 · Implement information extraction solutions 65

Recorded as checked against sources460 of 460

questions recorded as having their answer, options, and explanation checked against official Microsoft documentation

How questions are made →

Exam domains

What's on the exam

Implementing generative AI and agentic solutions carries the most weight at 30–35%, covering building Foundry-based generative apps, constructing agents with tool-calling and multi-agent orchestration, and operationalizing those systems in production. Planning and managing an AI solution follows at 25–30%: choosing the right Foundry service and model, setting up infrastructure and CI/CD, and securing and monitoring what's running. Computer vision, text and speech, and information extraction each sit at 10–15% — narrower but still tested in depth.

The split favors builders over integrators: more of the exam asks you to construct an agent or generative workflow than to wire up a single pre-built vision or speech call.

Exam blueprint: AI-103

Plan and manage an Azure AI solution25–30%

Choose appropriate Microsoft Foundry services and models, set up and deploy AI solutions, manage/monitor/secure AI systems, and implement responsible AI across generative and agentic systems.

≈ 12 h

Implement generative AI and agentic solutions30–35%

Build generative apps with Foundry (RAG, tool-augmented flows), build agents (roles, function-calling, memory, multi-agent orchestration), and optimize and operationalize generative AI systems.

≈ 15 h

Implement computer vision solutions10–15%

Image and video generation and editing, multimodal understanding and captioning, video analysis, and responsible AI for multimodal content.

≈ 6 h

Implement text analysis solutions10–15%

Language-model text analysis (entities, sentiment, translation, custom outputs) and speech solutions (speech-to-text, text-to-speech, speech as an agent modality).

≈ 6 h

Implement information extraction solutions10–15%

Build retrieval and grounding pipelines (semantic, hybrid, and vector search; RAG ingestion) and extract content from documents with Content Understanding.

≈ 6 h

Format

Exam format and question types

Expect roughly 40 to 60 questions in about 120 minutes. Item types include single-answer multiple choice, multiple-response (select all that apply), and drag-and-drop ordering or mapping, with scenario-based stems that ask you to apply two or more concepts together. The score is reported on a 1 to 1000 scale and 700 is required to pass. AI-103 is built on Microsoft Foundry. Most questions cover generally available features; preview features appear only where they are commonly used.

Question types: AI-103

Multiple Choice60%

Select the single answer that best meets the question’s requirements.

Multiple Response25%

Select multiple answers. Follow the question’s instructions on how many to choose.

Drag & Drop15%

Move items into the slots, groups, or sequence specified by the task.

See Microsoft for official question-format information. The shares shown describe the Pass-IT practice pool; they do not establish the proportions on the official exam.

Planning

Preparing for AI-103

AI-103 is delivered through Pearson VUE, either at a test center or with online proctoring, and is currently offered in English only. The certification is valid for one year, and Microsoft opens a free renewal assessment on Learn before it expires.

Preparation and logistics: AI-103

Preparation

Illustrative study time25–70 h

illustrative planning range: 25 h with relevant experience to 70 h when starting out; your needs may fall outside this range

LevelAssociate

Recommended backgroundExperience developing apps with Python; familiarity with general AI, generative AI, and Azure AI services.

Taking and maintaining the certification

DeliveryPearson VUE (test center or online proctored)

Retake policy24-hour wait after a failed first attempt; escalating waits after; max 5 attempts per 12 months

Certification validity1 year

Renew for free each year by passing the online renewal assessment on Microsoft Learn before the certification expires.

Official guideView on provider site

Study notes

Common pitfalls

Topics to review: AI-103

01Connected Agents vs Foundry workflows
Classic Connected Agents is deprecated and absent from the new Foundry Agent Service; orchestrate multiple agents with Foundry workflows, the Microsoft Agent Framework, or the A2A tool.
02RAG before fine-tuning
Ground responses with retrieval (Azure AI Search hybrid/vector search) rather than fine-tuning when the goal is adding domain knowledge or reducing fabrications.
03Keyless, managed-identity auth
Secure Foundry resources with managed identity, private networking, and keyless credentials instead of storing API keys.
04Content Understanding vs Document Intelligence
Use Azure AI Content Understanding for multimodal extraction and grounded RAG inputs; do not assume Document Intelligence covers video/audio.
05Responsible AI is tested
Know content filters, safety evaluators, risk detection, trace logging, and provenance metadata as first-class exam topics, not afterthoughts.
06GA over preview
Answer for generally available behaviour; preview features appear only when commonly used.

Practice with questions and explanations in Pass-IT →

Questions

Frequently asked questions

How long is the Developing AI Apps and Agents on Azure exam?

The AI-103 exam has 50 questions and a 120-minute time limit.

What is the passing score for Developing AI Apps and Agents on Azure?

The passing score for the AI-103 exam is 700 / 1000.

Which pitfalls should I review when preparing for Developing AI Apps and Agents on Azure?

Topics to review include Connected Agents vs Foundry workflows, RAG before fine-tuning, Keyless, managed-identity auth, Content Understanding vs Document Intelligence, Responsible AI is tested, GA over preview. Work through examples to check that you understand the distinctions and can explain your answer.

Is AI-103 the replacement for AI-102?

Yes. Microsoft retired AI-102 at the end of June 2026 and our catalog records AI-103 as its successor. The content moved with it: generative AI and agentic solutions is now the largest area at 33%, ground AI-102 barely touched. Our guide covers what changed and what to do if you were mid-preparation. Read the AI-102 to AI-103 guide

How is the AI-103 exam weighted?

Implementing generative AI and agentic solutions is the largest area at 33%, with planning and managing an Azure AI solution at 27%. Information extraction takes 14%, and computer vision and text analysis 13% each. Six tenths of the exam is generative AI and solution management, so the classic vision and language services are now the smaller half.

What do you need to know before AI-103?

Microsoft expects experience developing applications in Python and familiarity with AI, generative AI and the Azure AI services, without enforcing a certification prerequisite. The catalog budget is around 45 hours. It is a developer exam, so SDK behaviour matters more than machine learning theory.

How long does AI-103 stay valid?

One year. Microsoft renews it through a free online renewal assessment on Microsoft Learn that has to be passed before the expiry date. On a subject moving this quickly, the yearly assessment is a fair proxy for whether the knowledge is still current.

One certification, 12 months

Practice for AI-103

Focus your practice on one certification, or choose Pro to practice across all certifications.

Buy AI-103 access for $29.99

Start a free practice sessionTry the first 20 questions without a card to see whether the practice suits you.

For eligible purchases: money-back guarantee if you fail your exam.

View guarantee terms →