AI-103 replaced AI-102: what the Azure AI exam tests now

Microsoft retired exam AI-102 on 30 June 2026. Its replacement is exam AI-103, which leads to the Microsoft Certified: Azure AI Apps and Agents Developer Associate. It is still an intermediate exam for Azure AI engineers, at $165 and 120 minutes, but it tests a different job: building generative apps and agents on Microsoft Foundry, rather than calling Azure's separate AI services one by one. Generative AI and agents are now 30–35% of the exam, the guide shrank from 106 skill lines to 64, and it expects Python.

If you passed AI-102, your certification is still valid, but you can't renew it. More on that below.

What happens to your AI-102 certification

Microsoft's line is that "retirement does not revoke or invalidate Certifications that were earned". Your Azure AI Engineer Associate stays on your transcript until it expires.

The catch is the expiry. Microsoft's associate certifications last a year and are renewed with a free online assessment on Microsoft Learn, and a retiring certification can only be renewed before it retires. So the last AI-102 renewals happened before 30 June 2026. Each AI-102 certification now runs to its expiry date and stops. If your employer, a partner programme or a client counts it, plan AI-103 before that date.

If you were halfway through AI-102 study, Microsoft's advice for anyone "not close to testing" was to switch to AI-103, and much of what you learned carries over (see what carries over).

Where the weight moved

AI-102 had 6 sections; AI-103 has 5. Agents no longer have a small section of their own; they are joined with generative AI into the largest section:

Read the middle row first. On AI-102 agents were 5–10% of the exam. On AI-103, agents and generative apps together are the biggest share, and the planning section, now also bigger, is mostly about running them: quotas, cost, monitoring, security and governance of agents.

What's new in AI-103

Read section by section, the new lines describe someone who ships agents to production rather than someone who wires up an API:

• Agents as systems. Agent roles, goals and tool schemas; conversation memory and how the agent keeps track of a conversation; multi-agent orchestration; "autonomous or semiautonomous workflows with safeguards and approval flow controls"; and governing agents "with oversight modes, constraints, and tool-access controls".
• Running them. Managing "quotas, scaling, rate limits, and cost footprints" for model and agent workloads; monitoring drift, safety events and grounding quality; tracing, token analytics and latency breakdowns; auditing through trace logs, provenance metadata and approval workflows.
• Security without keys. Managed identity, private networking and "keyless credentials", where AI-102 asked you to "manage and protect account keys".
• Better answers. Detecting fabrications when evaluating models and apps, model reflection and self-critique loops, chain-of-thought evaluations, and combining LLMs with rules engines.
• Generation, not only analysis. Computer vision now opens with generating images and videos from prompts and reference media, inpainting and mask-based edits, and editing generated video. AI-102's only generation line was "Use the DALL-E model to generate images".
• Multimodal safety. Detecting indirect prompt injection hidden as text inside images, and enforcing visual policy rules such as watermarks and brand use.
• Small models. Choosing between large and small language models, multimodal models and Foundry Tools for each task.
• Content Understanding everywhere. Azure Content Understanding in Foundry Tools appears in vision (single-task and pro-mode pipelines) and in extraction (clean, grounded representations for agents and RAG). In AI-102 it was one line at the very end.

What AI-103 no longer names

The study guide says its bullets "illustrate" each skill and that "related topics may be covered", so a missing line is not a promise you won't see the topic. Still, these AI-102 tasks have no line of their own in AI-103:

• Language understanding by hand: creating intents and entities, adding utterances, training and backing up a language understanding model.
• Custom question answering: knowledge bases, question-and-answer pairs, chit-chat and multi-turn conversations.
• Custom Vision training: labelling images, training, evaluating and publishing an image classification or object detection model.
• Classic search plumbing: Knowledge Store projections, and query syntax with sorting, filters and wildcards. Search itself stays, as semantic, hybrid and vector search for grounding.
• Document Intelligence model building: custom, composed and trained models. AI-103 tests extraction through multimodal pipelines and Content Understanding instead.
• Older tools: prompt flow, containers on local and edge devices, Azure AI Video Indexer, Spatial Analysis, and SSML for speech.

The pattern is clear: wherever a task used to mean training a small dedicated model, AI-103 expects a prompt to a general model, or a Foundry tool, to do it.

What carries over from AI-102

Less is lost than the list above suggests. Planning and responsible AI (content filters, prompt shields, harm detection) carry over in new words. So do RAG, evaluation, prompt engineering, generation parameters, multi-model orchestration and CI/CD for Foundry projects. Image analysis, captions, object detection, OCR, translation, sentiment and speech-to-text all stay, now mostly through multimodal models. The speech lines now frame speech as part of an agent ("integrate speech as an agent modality"), custom speech models included.

Two changes need real study time. One is the agents section: if you prepared for AI-102 in 2025, its 5–10% probably got the least of your time, and it is now the core of the exam. The other is the language: AI-102 accepted Python or C#, and AI-103's guide asks for experience "developing apps by using Python". If you wrote C# answers, practise reading the Foundry Python SDK.

How to study for AI-103

Microsoft's own course is free on Microsoft Learn: AI-103T00, Develop AI apps and agents on Azure, in 4 learning paths.

That is about 30 hours of reading and labs. The agents path is the longest, as it should be given the weights. After it:

1. Build one agent end to end in Python on Foundry: retrieval over your own documents, one function tool, conversation memory, tracing turned on and a managed identity instead of a key. Most lines in the 2 largest sections touch a project like that. Our free AI engineering course builds the same kind of project step by step.
2. Take the free practice assessment on AI Skills Navigator (you need to sign in), and try the exam sandbox to see the question types. The exam is proctored and may include interactive parts.
3. Read the docs for preview features you use. Most questions are about generally available features, but the guide says preview features can appear "if those features are commonly used". Foundry's agent features change fast, so read the current docs, not a course from last year.

If you take the exam in a language it isn't offered in, you can ask for 30 extra minutes. If you fail, you can retake it after 24 hours.

Is it the right exam for you?

AI-103 is the Azure exam for building AI apps and agents. If your agents run on another cloud or on Databricks, our best AI certifications in 2026 lists the matching exam on each platform. If you build on Azure, expect the exam to assume Microsoft Foundry throughout.