Update (October 2026): This guide was corrected against Microsoft Learn on 20 October 2026. It now uses the five official skill areas and their weights, the published exam time of 120 minutes, the official certification name, the exact learning path and module names, and today's Microsoft Foundry terms. Earlier versions used a four-domain table written before Microsoft published the skills outline.
The AI-103 exam, "Developing AI Apps and Agents on Azure", earns the Microsoft Certified: Azure AI Apps and Agents Developer Associate certification. You have 120 minutes, you need a scaled score of 700 to pass, and you can open Microsoft Learn during the exam. It tests five skill areas, the biggest being generative AI and agents at 30 to 35%. The questions assume you build with Python and Microsoft Foundry. This guide covers the exam facts, the five skill areas, the four official learning paths module by module, the code patterns that matter, and a study plan.
What Is the AI-103 Exam?
AI-103 replaced AI-102, which retired on June 30, 2026, together with AI-900. It validates that you can design, build and deploy AI apps and agents on Azure using Microsoft Foundry (called Azure AI Foundry until its rename), Microsoft's platform for models, agents, tools, evaluation and safety.
The exam at a glance (all from Microsoft Learn):
| Fact | Detail |
|---|---|
| Exam | AI-103: Developing AI Apps and Agents on Azure |
| Certification | Microsoft Certified: Azure AI Apps and Agents Developer Associate |
| Time | 120 minutes; Microsoft lists 140 minutes of seat time for associate exams that may contain labs |
| Questions | Not published for AI-103; Microsoft says most exams have 40 to 60 |
| Pass mark | 700 (a scaled score, so not the same as 70%) |
| Interactive tasks | Possible: "You may have interactive components to complete as part of this exam" |
| Open book | Yes: Microsoft Learn is available during the exam, except Q&A, practice assessments and your profile, and no extra time is added |
| Language for code | Python |
| Prerequisites | None required |
| Booking | Through Pearson VUE, online or at a test centre, up to 90 days ahead |
| Exam languages | English, Chinese (Simplified and Traditional), French, German, Japanese, Korean, Italian, Portuguese (Brazil), Spanish |
| Practice | Free official practice assessment on AI Skills Navigator (sign-in needed) |
| Renewal | Yearly, through a free online assessment on Microsoft Learn |
| Price | Set by the country where you take it |
The official training course is AI-103T00-A: Develop AI apps and agents on Azure, a four-day course with four learning paths and 30 modules: 6, 9, 7 and 8 per path, with hands-on exercises throughout.
Key Differences: AI-102 vs AI-103
| Area | AI-102 (retired June 30, 2026) | AI-103 (current) |
|---|---|---|
| Certification | Azure AI Engineer Associate | Azure AI Apps and Agents Developer Associate |
| Platform | Azure AI Services (Cognitive Services) and Azure OpenAI | Microsoft Foundry: one resource with projects, models, agents and Foundry Tools |
| Centre of gravity | Calling individual AI service APIs | Building generative AI apps and agents end to end |
| Agents | A smaller part of the exam | Prompt and hosted agents, tools, MCP, Foundry IQ, multi-agent orchestration, A2A |
| Safety | Content filters and responsible AI principles | Guardrails on models and agents, evaluation, red teaming, monitoring |
| Speech | Speech SDK | Azure Speech in Foundry Tools, speech models, Voice Live |
| Vision | Computer Vision APIs | Multimodal models, image and video generation, Content Understanding |
If you hold AI-102, it stays on your transcript until it expires, but it can no longer be renewed: Microsoft retired both the certification and its renewal assessment.
Try a real Azure AI lab, free
A pre-provisioned Azure environment in your browser. No Azure account, no credit card, no setup.
Most large enterprises already run on Microsoft infrastructure. Azure AI Foundry is the platform those teams use to build and deploy AI apps, so the skills you practice here map directly to real job demand.
The Official AI-103 Curriculum: All Four Learning Paths
The course has four learning paths. The names below are Microsoft's own, with each module's unit count from Microsoft Learn.
Learning Path 1: Develop generative AI apps in Azure
The foundation: how Foundry is organised, how to choose and call models, and how to make answers better and safer.
Module 1: Plan and prepare to develop AI solutions on Azure (9 units)
The Foundry landscape: a Foundry resource with projects inside it, model deployments, connections, Foundry Tools (Language, Speech, Vision, Translator, Content Safety), developer tooling (VS Code, Azure CLI, SDKs) and responsible AI principles. Older material talks about hubs. Those still exist only in Foundry (classic); new work uses a Foundry resource with projects, and that is what the exam's infrastructure questions mean.
Module 2: Select, deploy, and evaluate Microsoft Foundry models (8 units)
The Foundry model catalog (OpenAI, Phi, Llama, Mistral and thousands more), deployment types such as Global Standard, Data Zone and provisioned throughput, benchmarks and the built-in evaluators. You are tested on choosing a model for a requirement, not on memorising the catalog.
Module 3: Develop a generative AI chat app with Microsoft Foundry (8 units)
Hands-on with the Foundry SDK and the Responses API, which Microsoft recommends for most new development. Topics: the project endpoint, keyless authentication with DefaultAzureCredential, multi-turn conversations and the response structure.
The current pattern, with azure-ai-projects 2.x and the openai package:
from azure.ai.projects import AIProjectClient
from azure.identity import DefaultAzureCredential
project = AIProjectClient(endpoint=PROJECT_ENDPOINT, credential=DefaultAzureCredential())
openai = project.get_openai_client()
response = openai.responses.create(model="gpt-5-mini", input="Summarise our refund policy in two sentences.")
print(response.output_text)
If a tutorial uses client.inference.get_azure_openai_client() or the azure-ai-inference package, it is out of date: that package was retired on August 26, 2026.
Module 4: Develop generative AI apps that use tools (9 units)
Tool calling (function calling): describing tools so the model can decide when to use them, handling the tool request, returning the result and chaining calls. A key exam concept: the model does not execute the tool. Your application does, then sends back the result.
Module 5: Optimize generative AI model performance with Microsoft Foundry (8 units)
Three strategies, and when each one fits:
- Prompt engineering: system messages, few-shot examples, step-by-step instructions, sampling settings.
- RAG (retrieval augmented generation): an Azure AI Search index, chunking, embedding models, grounding answers in your data.
- Fine-tuning: when it beats RAG, supervised fine-tuning with JSONL data, and evaluating the result.
Choosing between the three is a common question type.
Module 6: Implement a responsible generative AI solution in Microsoft Foundry (9 units)
Microsoft's four stages: Map, Measure, Mitigate, Manage. You map potential harms, measure them with evaluations and red teaming, mitigate them with guardrails (Microsoft's current term for content filters: controls applied to user input, tool calls, tool responses and output), and manage the solution in production. Responsible AI is not a separate exam domain, but it appears inside "Plan and manage an Azure AI solution" and again for multimodal content.
Learning Path 2: Develop AI agents on Azure
The largest path, and the heart of the exam's biggest skill area.
Module 1: Develop AI agents with Microsoft Foundry and Visual Studio Code (11 units)
Building agents in Foundry Agent Service from the portal, the VS Code extension and the SDK. Agents now run on the Responses API, so the terms are conversations, items, responses and agent versions. If you see threads, messages, runs and run steps, that is the classic agent model, which Microsoft has deprecated and will retire on March 31, 2027.
Module 2: Integrate custom tools into your agent (7 units)
Giving an agent your own functions as tools: defining them, attaching them, handling the calls, and testing the result.
Module 3: Integrate MCP Tools with Azure AI Agents (7 units)
Model Context Protocol (MCP), the open standard that lets an agent discover and call external tools at runtime: the host, client and server roles, and connecting an MCP server to a Foundry agent with the MCPTool class.
from azure.ai.projects.models import MCPTool
mcp_tool = MCPTool(
server_label="azure-language",
server_url="https://{resource}.cognitiveservices.azure.com/language/mcp?api-version=2025-11-15-preview",
require_approval="always",
)
In a full agent, this tool is attached when you create the agent version, and with require_approval="always" your code must answer each approval request before the call runs.
Module 4: Build knowledge-enhanced AI agents with Foundry IQ (8 units)
Foundry IQ, the managed knowledge layer for agents: knowledge bases built on Azure AI Search that give an agent your organisation's documents without custom retrieval code.
Module 5: Integrate your agent with Microsoft 365 (9 units)
Publishing a Foundry agent to Microsoft Teams and Microsoft 365 Copilot, publish scopes, and Work IQ for Microsoft 365 data. Know the difference between sharing for testing and publishing to the organisation.
Module 6: Build agent-driven workflows using Microsoft Foundry (11 units)
Workflows of steps, branches, loops and human checkpoints. Note the date: Microsoft is retiring the portal workflows on December 1, 2026, and points new workflows to Microsoft Agent Framework. Learn the concepts; build new ones in code.
Module 7: Develop an AI agent with Microsoft Agent Framework (7 units)
Microsoft Agent Framework, the code-first framework for agents. Microsoft describes it as the direct successor to Semantic Kernel and AutoGen, and it works with Foundry, Azure OpenAI and other providers.
Module 8: Orchestrate a multi-agent solution using the Microsoft Agent Framework (11 units)
The five orchestration patterns:
- Concurrent: agents work in parallel and the results are merged.
- Sequential: each agent passes its output to the next.
- Group chat: agents converse to reach an answer.
- Handoff: one agent hands control to a specialist.
- Magentic: a manager agent plans and directs the others.
Module 9: Discover Azure AI Agents with A2A (8 units)
Agent-to-Agent (A2A): Python clients that discover and call remote agents, including across organisational boundaries.
Learning Path 3: Develop natural language solutions in Azure
Text, speech and translation, both through Foundry Tools and through models.
Module 1: Analyze text with Azure Language in Foundry Tools (8 units)
The module's objectives are language detection, named entity recognition and PII extraction, the capabilities Microsoft now treats as core. Sentiment, key phrases, summarisation and entity linking still work but are marked deprecated, with retirement dates in 2028 and 2029.
from azure.ai.textanalytics import TextAnalyticsClient
from azure.identity import DefaultAzureCredential
client = TextAnalyticsClient(endpoint=LANGUAGE_ENDPOINT, credential=DefaultAzureCredential())
result = client.recognize_pii_entities(documents=["Call Ana on +31 6 1234 5678 about invoice 4471."])[0]
print(result.redacted_text)
Module 2: Develop a text analysis agent with the Azure Language MCP server (6 units)
Giving an agent Language capabilities through MCP. The module teaches four: language detection, named entity recognition, PII redaction and Text Analytics for health. The agent picks the right one from the user's request, with no routing code.
Module 3: Develop a speech-capable generative AI application (7 units)
Speech models: gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize for speech to text, and gpt-4o-tts and gpt-4o-mini-tts for text to speech.
from openai import AzureOpenAI
client = AzureOpenAI(azure_endpoint=FOUNDRY_ENDPOINT, api_key=FOUNDRY_KEY, api_version="2025-03-01-preview")
with open("audio.mp3", "rb") as f:
transcription = client.audio.transcriptions.create(model=MODEL_DEPLOYMENT, file=f, response_format="text")
Module 4: Create speech-enabled apps with Azure Speech in Microsoft Foundry Tools (9 units)
The Speech SDK: SpeechConfig, SpeechRecognizer and SpeechSynthesizer, audio output formats, neural voices, and SSML for speaking style, pauses, pronunciation and prosody.
Module 5: Develop a speech agent with the Azure Speech MCP server (6 units)
The Speech MCP server gives an agent speech to text and text to speech. It needs an Azure Blob Storage account for the audio files, connected with a SAS URL, which is an easy exam detail to miss.
Module 6: Develop an Azure Speech Voice Live Agent in Microsoft Foundry (7 units)
The Voice Live API: real-time, two-way spoken conversations over WebSocket, with voice activity detection, noise reduction, echo cancellation and avatars. The Python SDK is async only.
from azure.ai.voicelive.aio import connect
from azure.ai.voicelive.models import RequestSession, Modality, ServerVad
async with connect(endpoint=ENDPOINT, credential=AzureKeyCredential(API_KEY), model="gpt-4o") as conn:
session = RequestSession(
modalities=[Modality.TEXT, Modality.AUDIO],
instructions="You are a helpful assistant.",
turn_detection=ServerVad(threshold=0.5, silence_duration_ms=500),
)
await conn.session.update(session=session)
Module 7: Translate text and speech with Microsoft Foundry Tools (7 units)
Two services:
- Azure Translator in Foundry Tools: text and document translation and custom translation, with
TextTranslationClient. - Speech translation:
SpeechTranslationConfigwithTranslationRecognizerfor real-time spoken translation, using either manual synthesis (several languages) or event-based synthesis (one target language).
Learning Path 4: Extract insights from visual data on Azure
Images, video and documents.
Module 1: Develop a vision-enabled generative AI application (6 units)
Multimodal models in chat apps: sending images as base64 data URLs or web URLs with the prompt.
Module 2: Generate images with AI (7 units)
Image generation models from the catalog, trying them in the playground, and calling them from an app.
Module 3: Generate videos with Microsoft Foundry (7 units)
Video generation with Sora 2 from a text prompt or a reference image, in clips of 4, 8 or 12 seconds.
Module 4: Analyze images with Content Understanding (6 units)
Azure Content Understanding: analyzers that extract the fields you define from images, with confidence scores.
Module 5: Create a multimodal analysis solution with Azure Content Understanding (7 units)
Content Understanding across documents, audio and video, built visually in the portal.
Module 6: Create an Azure Content Understanding client application (7 units)
The same analyzers from code. The module's two objectives: use the REST API to build an analyzer, and use it to consume one. The workflow:
- Get the Foundry resource endpoint and credentials.
- Define the schema, from a sample and an analyzer template.
- Create the analyzer with an API call.
- Send new content and read the extracted fields in code.
Exam note: Modules 5 and 6 are the visual and the code routes to the same thing. Expect scenarios that ask which one fits a requirement.
Module 7: Extract data with Azure Document Intelligence (8 units)
Three model types:
- Prebuilt models: invoices, receipts, IDs, health insurance cards, US tax forms, bank statements, contracts, pay stubs and mortgage documents.
- Layout model: text, tables, barcodes, formulas and font information.
- Custom models: extractors trained on your own documents.
Module 8: Create a knowledge mining solution with Azure AI Search (9 units)
Azure AI Search as the retrieval layer: index design, skillsets with OCR and AI enrichment, indexers and queries. It connects back to RAG in Learning Path 1.
AI-103 Exam Domains and Weights, Mapped to Learning Paths
Microsoft's skills outline ("Skills measured as of April 16, 2026") has five areas:
| Skill area | Weight | Where it is taught |
|---|---|---|
| Plan and manage an Azure AI solution | 25 to 30% | LP1 Modules 1, 2 and 6; security, networking, monitoring and responsible AI across the course |
| Implement generative AI and agentic solutions | 30 to 35% | LP1 Modules 3 to 5; all of LP2 |
| Implement computer vision solutions | 10 to 15% | LP4 Modules 1 to 5 |
| Implement text analysis solutions | 10 to 15% | LP3 (language, speech, translation) |
| Implement information extraction solutions | 10 to 15% | LP4 Modules 4 to 8 (Content Understanding, Document Intelligence, Azure AI Search) |
The biggest area, generative AI and agentic solutions at 30 to 35%, covers chat apps, tools, RAG, agents and multi-agent orchestration. "Plan and manage" is the second biggest, and candidates underestimate it: it covers Foundry resource and project design, deployment choices, keyless credentials, private networking, roles, guardrails and monitoring.
Still reading? Grab the AI-103 exam study guide
One-page reference covering Azure AI Foundry services, endpoints, and key SDK calls.
Study Strategy
Phase 1: Architecture First (Days 1 to 3)
Before any module, draw Foundry from memory: Foundry resource, then projects, then deployments, connections and endpoints. Know the difference between the resource endpoint (used by Foundry Tools) and the project endpoint (used by agents and the Foundry SDK). Know when to use DefaultAzureCredential rather than a key, and which role a caller needs.
Phase 2: Learning Paths 1 and 2 (Days 4 to 14)
Work through LP1 and LP2 in order. The agent modules assume you can already build a chat app and use tool calling. Do each module's exercise in a real Azure environment.
Build the LP1 Module 5 RAG pipeline end to end as one exercise:
- Create an Azure AI Search index over your own documents.
- Choose a chunking strategy and an embedding model.
- Connect the index to a model call in Foundry.
- Generate a grounded answer.
- Score it with the built-in evaluators.
That one exercise touches three of the five skill areas.
Phase 3: Learning Paths 3 and 4 (Days 15 to 20)
The language and speech modules are SDK-heavy. Know which client does what: TextAnalyticsClient for language, SpeechConfig with SpeechRecognizer and SpeechSynthesizer for classic speech, the OpenAI client for the speech models, and Voice Live for real-time conversation.
For documents, know when a Document Intelligence prebuilt model fits and when Content Understanding is the better choice. The Foundry Tools guide compares them service by service.
Phase 4: Responsible AI and Weak Spots (Days 21 to 25)
Finish LP1 Module 6 and take the free official practice assessment on AI Skills Navigator. Revisit any area where you score below 70%. Know the four stages (Map, Measure, Mitigate, Manage), the difference between guardrails and evaluation, and when a human must review an output.
Key Lab Exercises to Prioritise
| Exercise | Skill areas |
|---|---|
| Build a generative AI chat app | Generative AI and agentic |
| Build an agent with custom tools | Generative AI and agentic |
| Connect MCP tools to an agent | Generative AI and agentic; text analysis |
| Implement a RAG pipeline | Generative AI and agentic; information extraction |
| Analyze text with Azure Language | Text analysis |
| Build a Content Understanding client application | Information extraction; computer vision |
| Build a text analysis agent with the Language MCP server | Text analysis |
| Build a speech-enabled app | Text analysis |
| Build a vision-enabled app | Computer vision |
| Extract data with Document Intelligence | Information extraction |
| Configure guardrails and evaluations | Plan and manage |
Frequently Asked Questions
How long is the AI-103 exam? 120 minutes. Microsoft lists 140 minutes of seat time for associate exams that may contain labs, to cover the agreement and the survey.
How many questions are on the AI-103 exam? Microsoft does not publish a number for AI-103. It says most Microsoft Certification exams have between 40 and 60 questions, and the number varies.
What is the passing score for AI-103? 700 or higher. It is a scaled score, so it does not mean exactly 70% of the points.
Is the AI-103 exam open book? Partly. You can use Microsoft Learn during the exam, except Q&A, practice assessments and your profile. No extra time is added, so treat it as a way to check a detail, not to learn a topic.
Do I need Python for AI-103? Yes. Microsoft says candidates should have experience developing apps with Python, and the certification validates skills "using Python and Microsoft Foundry". You need to read and complete code patterns, especially which client or SDK call matches which capability.
Are there labs on the AI-103 exam? There may be. Microsoft says you may have interactive components to complete, and it does not publish which exams include labs. Hands-on practice is the best preparation either way.
Are there prerequisites for AI-103? No formal prerequisite. Microsoft recommends experience building apps with Python and familiarity with AI concepts and Azure. If you are new to Azure AI, AI-901 is the gentler start.
How do I schedule the AI-103 exam? From the exam page on Microsoft Learn, through Pearson VUE, either online or at a test centre, up to 90 days in advance. The price depends on the country where you take it.
Is there a free AI-103 practice test? Yes. Microsoft's official practice assessment is free on AI Skills Navigator, where you need to sign in.
If I already have AI-102, do I need to take AI-103? Your AI-102 certification stays on your transcript until it expires, but it cannot be renewed because Microsoft retired it on June 30, 2026. To keep a current Azure AI developer certification, you need AI-103.
What comes after AI-103? The expert level: Microsoft Certified: Multi-Agent AI Solutions Expert, through exam AI-500, accepts AI-103 as a prerequisite.
How do I renew the AI-103 certification? Every year, with a free online renewal assessment on Microsoft Learn, which opens six months before the certification expires.
What Azure services appear on the AI-103 exam? Microsoft Foundry (resource, projects, models, Agent Service, Foundry IQ, evaluations and guardrails), Azure OpenAI models, Azure Language, Azure Speech and Azure Translator in Foundry Tools, Azure Content Understanding, Azure Document Intelligence, Azure AI Search, Content Safety and Azure Blob Storage.
What is MCP and why is it on the AI-103 exam? MCP (Model Context Protocol) is an open standard that lets agents discover and call external tools at runtime. AI-103 uses it to connect Azure Language, Azure Speech and other services to Foundry agents, so expect questions on connecting MCP servers and on approval settings.
Timeline and Next Steps
AI-103 and AI-901 are the live Azure AI exams, and AI-102 and AI-900 retired on June 30, 2026. If you studied for AI-102, the Azure AI basics carry over, but plan time for Microsoft Foundry, agents and MCP.
For experienced Azure AI developers: plan 4 to 5 weeks, with hands-on exercises every week.
For candidates newer to Azure AI: take AI-901 first, then plan 6 to 8 weeks for AI-103.
For a module-by-module read of the exam content, the free AI-103 study guide on this site follows the same skill areas.
The official sources to bookmark: the AI-103 exam and certification page, the AI-103 study guide with the skills measured, and the AI-103T00 course.
The Hands-On Advantage
The official course teaches with exercises, concept by concept. That covers the exam. It does not always cover the job.
The GetHandsOn.ai AI-103 labs are scenario-based: each one is set inside a realistic business context, so you solve a problem the way a client would frame it rather than calling an API in isolation. That combination of exam readiness and practical readiness is what makes the certification credible to an employer.
What changed since this guide was first published
AI-900 and AI-102 retired on June 30, 2026, and AI-901 and AI-103 are now the live exams. Since then, three things changed that matter for your preparation:
- The platform terms. Azure AI Foundry is now Microsoft Foundry, hubs gave way to a Foundry resource with projects, Azure AI Services became Foundry Tools, and agents moved from threads and runs to the Responses API.
- The skills outline. Microsoft's April 16, 2026 outline has five skill areas, with "Plan and manage" at 25 to 30%.
- Retirements on the horizon. Classic agents retire on March 31, 2027, the portal workflow designer on December 1, 2026, and several Language features in 2028 and 2029.
The objectives that cost candidates the most marks
Four areas stand out. None of them are hard to read about. All of them are hard to answer from reading alone.
Evaluation. Candidates can define groundedness but cannot say what it does not measure. Groundedness is a precision check: it asks whether anything was made up. It says nothing about what the answer left out. If a question describes an agent that is accurate but incomplete, groundedness is the wrong evaluator. How to evaluate an AI agent in Microsoft Foundry walks through which evaluator answers which symptom.
Tool calling. Expect scenarios where the model picks the right intent and the wrong tool, or passes a malformed argument. Knowing the difference between tool selection, tool input accuracy and tool call success is worth more marks than knowing the SDK class names.
Retrieval design. RAG appears in the objectives, but several scenarios describe problems RAG cannot solve. Recognising when the answer is a database query, a deterministic rule or a fine-tuned model is a tested skill. When RAG is not the right answer covers the most common cases.
Multi-agent orchestration. Candidates who have only built single agents struggle here, because the questions are about handoff, shared state and containing failures rather than about prompts. Multi-agent orchestration on Azure covers the patterns the exam expects.
If Foundry itself is still unfamiliar, Microsoft Foundry explained covers the platform, and what Azure AI Foundry is used for covers the builds these objectives are written around.
If you want to work through these in a real Azure environment rather than on paper, the AI-103 course has 14 live labs.
Models are deployed when you need them. No Azure subscription needed. One-time payment, three months of access.
Related: What Is an AI Engineer in 2026? Role, Skills & Path