AI-102, the exam behind the Azure AI Engineer Associate certification, retired on June 30, 2026, and it can no longer be taken. AI-103, Developing AI Apps and Agents on Azure, is the exam Microsoft now offers for engineers who build AI apps and agents on Azure, and it earns the Azure AI Apps and Agents Developer Associate certification.
On paper the two look close. Both cover vision, language, speech, search and document extraction, and the final AI-102 outline already used Microsoft Foundry names and included Content Understanding. The difference is the work the questions describe. AI-102 spent much of its outline on training and calling individual AI services. AI-103 asks you to build apps and agents on Foundry that use those services as tools, and to govern, evaluate and monitor what you ship.
Find out where you stand first
If you studied for AI-102, take the free AI-103 practice questions before anything else. The questions you miss will show which of the changes below apply to you. The AI-103 course and six timed AI-103 practice exams on preporato.com are built on the current skills list.
The same jobs, then and now
Comparing the two skills lists job by job shows where the answers moved. The left column is what the AI-102 outline taught as of its December 23, 2025 version. The right column is the AI-103 outline as of April 16, 2026.
How each exam expects the job to be done
| Job | AI-102 skills list | AI-103 skills list |
|---|---|---|
| Answer questions from a policy library | A custom question answering project with question and answer pairs | An agent grounded through retrieval from Azure AI Search or a knowledge base |
| Pull structured details from a request | A language understanding model with intents, entities and utterances | Generative prompting or Foundry Tools that return entities and structured JSON |
| Find objects in product photos | A Custom Vision model you label, train and publish | Multimodal models and Content Understanding that identify objects and regions |
| Extract invoice fields | Document Intelligence prebuilt, custom and composed models | Content Understanding analyzers with structured or Markdown output |
| Create images | The DALL-E model | Image and video generation from text and reference media, with mask edits |
| Keep output safe | Content filters, blocklists and Prompt Shields | Guardrails, evaluators, trace logging, approval workflows and agent oversight |
| Build an agent | Foundry Agent Service and Agent Framework, a small slice of the outline | Agents at the center: tools, memory, orchestration, approvals and monitoring |
| Run AI at the edge | Container deployments for local and edge devices | No longer in the skills list |
Some of the AI-102 tools are on their way out of the product as well. Microsoft has set March 31, 2029 as the retirement date for conversational language understanding, custom question answering, custom text classification, key phrase extraction, sentiment analysis and summarization in Azure Language, and it directs new projects in those areas to Foundry models. DALL-E 3 retired on March 4, 2026.
Preparing for AI-103? Practice with 390+ exam questions
What the change looks like in a question
The table is easier to remember through scenarios. Here are three jobs an AI-102 candidate would recognize, solved the way AI-103 grades them.
A support assistant over a policy library
An insurer wants an assistant that answers customer questions from thousands of policy documents, shows where each answer came from, and opens a claim only after a staff member confirms it.
The AI-103 answer indexes the documents in Azure AI Search, with hybrid queries so exact policy numbers and paraphrased questions both match. A Foundry agent reaches that index through retrieval, either the Azure AI Search tool or a knowledge base in Foundry IQ, and the claim action becomes a function or MCP tool with require_approval set, so the agent pauses until a person decides. Groundedness and Response Completeness evaluations then show whether answers stay in the sources and cover what they should.
An AI-102 study plan would have reached for a custom question answering project. That knowledge base answers from curated question and answer pairs, and it has nothing to say about approvals or agent tools.
Invoices, contracts and amendments
A finance team processes supplier invoices, along with contracts whose prices change through amendments filed in the same PDF. It needs invoice fields in its ERP and the unit price that applies today.
In AI-103 terms, the invoices go through a copy of the prebuilt invoice analyzer saved under the team's own ID, so a change to the prebuilt definition can't break production. A classifier with enableSegment routes mixed files to the right analyzer. The current price comes from agentic mode in Content Understanding, a preview that reasons and calculates across one file and replaced the retired pro mode.
AI-102 covered extraction with Document Intelligence prebuilt and composed models, which return the values printed on the page. Working out a price from a base contract and three amendments was outside that scope.
Product images and a short clip
A marketing team needs product photos with new backgrounds and a short video for social media.
AI-103 expects the image edit API for a GPT-image model, with a PNG mask that leaves only the background editable and input_fidelity set to high to protect the product and logo. Sora 2 handles the video, and Content Credentials on the generated images record that they came from an Azure OpenAI model. AI-102 named the DALL-E model for image generation, and that model has since retired.
What carries over
Plenty of AI-102 preparation still counts. Azure AI Search indexing, enrichment skills, custom skills, and semantic and vector search sit in both outlines. AI-103 adds hybrid search and the connection from retrieval pipelines to workflows and agent tools, which is where agentic retrieval fits. Authentication and key management appear in both, and AI-103 spells out managed identity, private networking, keyless credentials and role policies. Speech to text, text to speech, custom speech, translation and PII detection carry over too.
The risk lies in the details around them. Service names, the resource model and the agent API have all moved, so a correct AI-102 habit can still pick the wrong option on AI-103.
Master These Concepts with Practice
Our AI-103 practice bundle includes:
- 6 full practice exams (390+ questions)
- Detailed explanations for every answer
- Domain-by-domain performance tracking
30-day money-back guarantee
If you studied for AI-102
Relearn these in order, since each one builds on the previous:
- The Foundry resource model. A Foundry resource with projects, model deployments and their types, quotas, and spillover from provisioned capacity.
- Agents on the Responses API. Agent versions, conversation state, tools through functions, OpenAPI and MCP, memory, and approval controls.
- Guardrails and evaluation. Deployment guardrails and their controls, Prompt Shields for document attacks, quality and agent evaluators, and red teaming.
- Multi-agent orchestration. The sequential, concurrent, handoff, group chat and Magentic patterns in Microsoft Agent Framework.
- Content Understanding. Prebuilt and custom analyzers, classifiers and routing, and agentic mode.
- Generation and multimodal safety. Image and video generation and editing, provenance, and injection through text in images.
The AI-103 course follows the exam's own order, and each lesson ends with an exam tip and checkpoint questions.
If you hold the AI-102 certification
Microsoft's retirement page says an earned certification stays on your transcript in your Microsoft Learn profile. It also says renewal is no longer an option once a certification retires. For Microsoft Partners, certifications earned before the retirement keep counting toward partner requirements for one year after the retirement date.
For a current credential in the same role, AI-103 is the exam to take, and your AI-102 experience shortens the list above to the parts that changed.
Frequently asked questions
Next step
The AI-103 complete guide covers the five skill areas and who the exam suits, and the AI-103 practice questions work through ten scenarios with every option explained. To go through the current objectives one by one, read AI-103 exam topics explained. When you're ready to rehearse, the six timed AI-103 practice exams score each skill area separately, and how to pass AI-103 on your first attempt covers the exam day itself.
Sources:
- Study guide for Exam AI-103: Developing AI Apps and Agents on Azure
- Study guide for Exam AI-102 (retired)
- Exam and assessment lab retirement
- What is Azure Language?
- Agentic mode in Content Understanding
- Image generation and editing in Microsoft Foundry
- Microsoft Agent Framework orchestrations
- What is Microsoft Foundry?
Ready to Pass the AI-103 Exam?
Join thousands who passed with Preporato practice tests