You define a function tool named get_stock_level with sku and warehouse parameters. In testing, the model sometimes omits warehouse and sometimes adds parameters that do not exist. You need the arguments to conform to the schema. What should you do?

Add warehouse to required, set additionalProperties to false and strict to true

Lower the temperature of the model deployment used by the agent

Set tool_choice to required so the function is always called

Describe the two parameters again in the agent instructions

A maintenance copilot uses vector-only queries against an Azure AI Search index. Engineers report that questions containing exact fault codes such as E4012 return loosely related chunks. Paraphrased questions work well. You need to improve results for both types of question. What should you do?

Add the question as search text alongside the vector query to run a hybrid query

Set exhaustive to true on the vector query so every vector is compared

Lower k on the vector query so fewer weak neighbours are returned

Change the similarity metric of the vector field from cosine to dotProduct

A dating app must screen profile photos for sexual and violent content at upload, before the photos are stored. No language model is involved in the upload flow. Which solution meets the requirement with the least development effort?

Call the Content Safety Analyze image API and apply thresholds in code

Assign a custom guardrail with High blocking levels to a chat deployment

Send each photo to a vision model and ask it whether the photo is safe

Create a custom Content Understanding analyzer with a classify field

Is AI-103 the replacement for AI-102?

Yes. Microsoft retired Exam AI-102 and the Azure AI Engineer Associate on 30 June 2026 and replaced them with Exam AI-103 and the Azure AI Apps and Agents Developer Associate. If you already hold the old certification, it stays valid until it expires.

How much does it cost and how long is it valid?

US$165 at list price, adjusted for the country where you sit it. Like other Microsoft associate certifications it lasts a year, and you renew it free each year with an online assessment on Microsoft Learn.

Do I need AI-901 Azure AI Fundamentals first?

No, it is not required. AI-901 helps if you are new to AI on Azure, but developers who already build with Azure services usually go straight to AI-103.

Which programming language does the exam use?

Python. Microsoft's audience profile says candidates should have experience developing apps with Python, so code in questions and labs is Python.

How long should I study?

Eight to twelve weeks part-time for most people who already build on Azure, longer if Microsoft Foundry, Azure AI Search and agents are new to you.

All certifications

Microsoft · AI-103 · Associate

Microsoft Certified: Azure AI Apps and Agents Developer Associate

Proves you can plan, build, secure and run generative AI apps and agents on Azure with Microsoft Foundry, from RAG and multi-agent workflows to vision, speech and document extraction.

Free practice: 216 questions written to the official exam guide, with the diagnostic, practice papers, timed mocks, a study plan and an explanation on every answer.

Official exam page

Find out where you stand on AI-103

20 questions drawn in the exam’s own domain weights. You get a report straight away: your score against the pass mark, where the points went, and what to study first.

Questions: 20
Time: about 20 minutes
Cost: free

Your email

Your answers save as you go, and the report stays in your record. A new email needs no code.

· your answers stay on this device and you can save the report at the end.

The course

The diagnostic shows where you stand. Then you learn the syllabus chapter by chapter, answer that chapter’s questions, and move on. When every chapter is mastered, you sit full timed mocks in the real exam’s shape.

15 chapters, about 11.5 hours in all. Each is a 45-minute study session with quick checks, then a 10-question quiz on that chapter.

Plan and manage an Azure AI solution28% of the exam

Choosing Foundry models and services45 min
Setting up and deploying AI solutions in Foundry45 min
Managing, monitoring and securing AI systems45 min
Responsible AI for generative and agentic systems45 min

Implement generative AI and agentic solutions32% of the exam

Building generative AI applications with Foundry45 min
Workflows, multistep pipelines and evaluation45 min
Building agents with Foundry Agent Service45 min
Multi-agent orchestration, autonomy and agent monitoring45 min
Optimising and operating generative AI systems45 min

Implement computer vision solutions13% of the exam

Generating and editing images and video45 min
Multimodal understanding and visual responsible AI45 min

Implement text analysis solutions13% of the exam

Text analysis and translation with language models and Foundry Tools45 min
Speech solutions for apps and agents45 min

Implement information extraction solutions14% of the exam

Retrieval and grounding pipelines with Azure AI Search45 min
Extracting content from documents with Content Understanding45 min

About the Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) exam

Microsoft Certified: Azure AI Apps and Agents Developer Associate is the Microsoft certification for engineers who build generative AI apps and agents on Azure. Exam AI-103 tests whether you can choose the right models and services in Microsoft Foundry, set them up securely, build retrieval-augmented and agentic solutions, and add vision, speech, text analysis and document extraction where the work needs them.

It replaced the Azure AI Engineer Associate. Microsoft retired Exam AI-102 on 30 June 2026, and AI-103 is now the exam for that role. The skills measured as of 16 April 2026 put the weight on generative AI and agents, at 30 to 35 percent of the exam, and on planning, securing and governing AI solutions, at 25 to 30 percent. Classic tasks such as custom vision training and question answering projects are gone; image and video generation, multimodal understanding with Azure Content Understanding in Foundry Tools, and multi-agent orchestration are in.

Microsoft describes the candidate as an Azure AI engineer who develops apps in Python and knows generative AI and Azure services. Nothing is formally required. Most people who already build on Azure plan on two to three months of part-time study.

The exam at a glance

Questions: Not published by Microsoft; Microsoft exams usually have 40 to 60 questions, and this one may include interactive components
Time: 120 minutes
Pass mark: 700 on a scale of 1 to 1,000
Exam fee: US$165 list price; Microsoft prices exams by the country you sit them in
Valid for: 1 year; renew free each year with an online assessment on Microsoft Learn
Where you sit it: Pearson VUE test centre or online proctored
Prerequisites: None required; Microsoft expects experience building apps in Python with generative AI and Azure services

As published in the vendor’s exam guide at the time of writing. Fees vary by country and change; the vendor’s page is the source of truth.

What is on the exam, domain by domain

Plan and manage an Azure AI solution
Choosing models and Foundry services for each task, setting up projects and deployments, CI/CD, quotas and cost, monitoring and security with managed identity and private networking, and responsible AI: guardrails, safety evaluations, auditing and agent governance.
28% of the paper
Implement generative AI and agentic solutions
The largest domain. Deploying and consuming models, RAG, workflows and multistep pipelines, evaluation, building agents with tools, retrieval and memory in Foundry Agent Service, multi-agent orchestration with approvals, and tracing and tuning in production.
32% of the paper
Implement computer vision solutions
Generating and editing images and video, multimodal understanding and captioning, visual question answering, Content Understanding for images and video, and responsible AI for visual content, including prompt injection hidden in images.
13% of the paper
Implement text analysis solutions
Extracting entities, summaries and structured JSON with language models and Foundry Tools, sentiment and safety detection, translation with Azure Translator or LLM flows, and speech to text, text to speech and speech translation for agents.
13% of the paper
Implement information extraction solutions
Ingesting and indexing content in Azure AI Search, semantic, hybrid and vector search, enrichment skills and OCR for RAG ingestion, and extracting structured or markdown output from documents with Content Understanding analyzers.
14% of the paper

How to prepare

Read the AI-103 study guide and study to its bullets. Much of the older AI-102 material, such as custom vision training and question answering projects, is no longer on the exam.
Build in a Microsoft Foundry project: deploy a chat model, add retrieval over your own documents with Azure AI Search, and call it from Python with the Foundry SDK and a managed identity.
Build an agent in Foundry Agent Service with a function tool and a retrieval tool, then trace it and run an evaluation. Agents and their oversight are a large share of the exam.
Learn the trade-offs that decide most questions: which model or Foundry Tool fits a task, keyword against semantic against vector against hybrid search, and keys against keyless authentication.
Take Microsoft's free Practice Assessment on AI Skills Navigator, then two timed practice papers above 80 percent before you book.

Sample questions

Three questions in the exam’s style, written by Square 1 against the published domains. They are ours, not the vendor's. The answer and the reasoning are under each one; the practice papers on this page draw from a bank written the same way, with an explanation on every question.

Question 1 · Implement generative AI and agentic solutions
You define a function tool named get_stock_level with sku and warehouse parameters. In testing, the model sometimes omits warehouse and sometimes adds parameters that do not exist. You need the arguments to conform to the schema. What should you do?
1. Lower the temperature of the model deployment used by the agent
2. Add warehouse to required, set additionalProperties to false and strict to true
3. Set tool_choice to required so the function is always called
4. Describe the two parameters again in the agent instructions
Answer: B. The required array stops fields being left out, and additionalProperties false with strict true asks for arguments that match the schema exactly. tool_choice required forces a call but does not control its arguments, and instructions or temperature changes do not enforce a schema.
Question 2 · Implement information extraction solutions
A maintenance copilot uses vector-only queries against an Azure AI Search index. Engineers report that questions containing exact fault codes such as E4012 return loosely related chunks. Paraphrased questions work well. You need to improve results for both types of question. What should you do?
1. Set exhaustive to true on the vector query so every vector is compared
2. Lower k on the vector query so fewer weak neighbours are returned
3. Change the similarity metric of the vector field from cosine to dotProduct
4. Add the question as search text alongside the vector query to run a hybrid query
Answer: D. Hybrid search adds a full-text query that matches exact strings such as codes, and RRF merges it with the vector results that handle paraphrase. Exhaustive KNN only improves recall of nearest neighbours; it still compares meaning rather than matching the exact code.
Question 3 · Implement computer vision solutions
A dating app must screen profile photos for sexual and violent content at upload, before the photos are stored. No language model is involved in the upload flow. Which solution meets the requirement with the least development effort?
1. Assign a custom guardrail with High blocking levels to a chat deployment
2. Send each photo to a vision model and ask it whether the photo is safe
3. Create a custom Content Understanding analyzer with a classify field
4. Call the Content Safety Analyze image API and apply thresholds in code
Answer: D. The Analyze image API in Azure AI Content Safety scores images for Hate, SelfHarm, Sexual and Violence on severities 0, 2, 4 and 6, and your code decides what to reject. A guardrail only scans traffic that flows through a model deployment, and this upload flow never calls a model.

Questions people ask before they book

Is AI-103 the replacement for AI-102?: Yes. Microsoft retired Exam AI-102 and the Azure AI Engineer Associate on 30 June 2026 and replaced them with Exam AI-103 and the Azure AI Apps and Agents Developer Associate. If you already hold the old certification, it stays valid until it expires.
How much does it cost and how long is it valid?: US$165 at list price, adjusted for the country where you sit it. Like other Microsoft associate certifications it lasts a year, and you renew it free each year with an online assessment on Microsoft Learn.
Do I need AI-901 Azure AI Fundamentals first?: No, it is not required. AI-901 helps if you are new to AI on Azure, but developers who already build with Azure services usually go straight to AI-103.
Which programming language does the exam use?: Python. Microsoft's audience profile says candidates should have experience developing apps with Python, so code in questions and labs is Python.
How long should I study?: Eight to twelve weeks part-time for most people who already build on Azure, longer if Microsoft Foundry, Azure AI Search and agents are new to you.