AI-102 Is Retired: How to Carry Azure AI Skills Into AI-103

AI-102 is no longer a current Microsoft exam. Microsoft retired the Azure AI Engineer Associate exam on June 30, 2026, and the modern Azure AI engineering path now centers on AI-103, Developing AI Apps and Agents on Azure.

That does not make everything learned for AI-102 obsolete. Computer vision, text analysis, search, document extraction, responsible AI, security, and production operations still matter. What changed is the center of gravity. The current Azure AI Apps and Agents Developer Associate role places much more emphasis on Microsoft Foundry, generative AI, agents, retrieval, multimodal applications, tool use, evaluation, observability, and governed deployment.

If you still have an old AI-102 course, lab repository, or study plan, the right response is not to throw it away or keep studying it unchanged. Treat it as a source of transferable engineering skills, then rebuild the plan around the current AI-103 blueprint.

Separate durable AI engineering from retired exam structure

AI-102 covered a broad Azure AI engineering role. Some of its service names and exam boundaries have changed, but the underlying questions remain familiar: which capability fits the requirement, how should it be secured, how is data processed, how is quality evaluated, and how does the solution behave after deployment?

Those questions are still useful. The problem is using the retired exam as the organizing framework. An old chapter sequence can overemphasize product patterns that are no longer central while underpreparing you for agents, generative workflows, Foundry projects, retrieval pipelines, and multimodal systems.

The old AI-102 exam is therefore best treated as historical context. It can help diagnose concepts you already know, but it should not define the final preparation checklist for a 2026 candidate.

AI-103 makes generative AI and agents a core engineering responsibility

Microsoft’s current AI-103 outline gives generative AI and agentic solutions the largest single share of the exam. Candidates are expected to build applications that use models, retrieval, tools, memory, workflows, evaluation, and agent orchestration rather than simply call a prebuilt cognitive endpoint.

That shift changes how you should practice. A useful lab should not end after sending a prompt to a model. Add retrieval from governed content. Add a tool with a narrow contract. Add an approval boundary. Trace the execution. Measure output quality, grounding, safety, and latency. Break one dependency and observe the failure path.

The AI-102 to AI-103 transition becomes clearer when old knowledge is mapped to current responsibilities rather than compared product name by product name.

Foundry knowledge now matters at architecture and operations level

AI-103 expects candidates to work with Microsoft Foundry as an environment for building and operating AI applications and agents. That includes selecting models and services, configuring deployments, integrating projects with development workflows, managing cost and quotas, and applying identity and network controls.

This is more than knowing where a button is located. You should understand how a Foundry project relates to an application, how the application authenticates, how models and agents are deployed, where retrieval data comes from, and what telemetry is available when something goes wrong.

Security deserves hands-on attention. Managed identity, role policies, private networking, keyless credentials, data boundaries, and tool permissions become more important when an AI application can retrieve sensitive content or invoke actions. A model response can be correct while the architecture remains unsafe.

Retrieval is now part of application behavior, not an optional search topic

Azure AI Search was already relevant to AI-102, but the current role connects retrieval much more directly to generative applications and agents. Candidates need to reason about ingestion, indexing, semantic and vector search, hybrid retrieval, enrichment, grounding, and how retrieved evidence flows into an application.

A good retrieval lab should include imperfect conditions. Put current and obsolete documents in the source set. Include overlapping terminology. Test an unauthorized user. Change one document and measure how quickly the index reflects it. Ask a question with no authoritative answer and see whether the application admits uncertainty or fabricates a response.

This is where older AI-102 search knowledge remains valuable. Indexes, data sources, enrichment, relevance, and query behavior still matter. The new expectation is to connect them to grounded generation and agent tools rather than treating search as an isolated workload.

Vision, text, speech, and information extraction still matter

AI-103 has not become an agents-only exam. Computer vision, text analysis, speech, and information extraction remain explicit parts of the role. What has changed is how naturally these capabilities combine with generative models and multimodal workflows.

For vision, practice interpreting images and video, producing accessible descriptions, and handling unsafe visual content. For text, practice entities, summarization, tone, translation, and structured output. For speech, connect speech recognition or synthesis to a wider conversational workflow. For document extraction, combine OCR, layout understanding, field extraction, and downstream reasoning.

The transferable lesson from AI-102 is to choose the capability based on the required input and output. The current extension is to ask how that capability participates in a larger AI application and which responsible-AI controls belong around it.

Evaluation and observability should be built into every lab

Older preparation often treated a successful API response as proof that the exercise worked. That standard is too low for current generative systems. AI-103 expects attention to model quality, grounding, safety, monitoring, traces, latency, and operational behavior.

Build small evaluation sets before changing prompts or models. Measure whether the application selects the right evidence, returns useful answers, produces valid structured output, uses tools correctly, and refuses or escalates when appropriate. After a change, rerun the same cases.

For agents, observability should include the chain of action: what the user asked, what evidence was retrieved, which tool was selected, which arguments were passed, what the tool returned, and whether a human approval was required. This makes troubleshooting much more rigorous than reading the final chat response.

Use old AI-102 material as a diagnostic, not as a syllabus

Keep an AI-102 resource when it teaches a durable skill well. Retire it when it depends on obsolete service behavior or when it ignores the current Foundry and agentic model.

A practical migration method is to create three columns: “still directly useful,” “useful but needs current implementation,” and “retired or low priority.” Computer vision concepts may remain directly useful. Search concepts may remain useful but need to be connected to modern RAG. An old Bot Framework workflow may be useful historically but insufficient for the current agentic objectives.

AI-103 transition readiness should separate genuinely transferable skills from areas that need current implementation practice. The point is not to maximize reuse; it is to avoid relearning durable fundamentals while still recognizing where the role changed.

Build one current end-to-end AI application

The most efficient preparation artifact is a small but complete application. Give it a real business requirement. Use a Foundry project. Choose a model deliberately. Add retrieval or multimodal input where the use case needs it. Add at least one governed tool. Apply identity and security. Add evaluation cases and traces. Document the failure modes.

Then extend the same application instead of creating ten disconnected demos. Add an agent. Add a new modality. Change the retrieval strategy. Introduce a safety rule. Add a human approval. This makes the relationships between the exam domains visible and gives you a better test of whether you can operate the solution rather than only configure a feature.

When you review mistakes, distinguish a product-knowledge gap from an architecture gap. Missing the name of a setting is different from misunderstanding why a managed identity is safer, why a RAG result is ungrounded, or why an autonomous tool needs an approval boundary.

Do not overcorrect by studying only agents and generative AI. Current AI engineering still depends on disciplined service selection, data handling, authentication, networking, monitoring, and failure recovery. A multimodal or agentic application can still fail because a storage permission, search index, endpoint configuration, quota, or network rule is wrong. The new exam changes emphasis; it does not remove engineering fundamentals.

A useful transition table has four columns: old AI-102 topic, current AI-103 responsibility, practical exercise, and evidence of readiness. For search, the exercise might be building hybrid retrieval and checking grounding. For vision, it might be processing several image types and handling an unsupported case. For agents, it might be adding one read-only tool, tracing the call, then introducing an authorization failure.

Review cost and lifecycle as well. Model deployments have quotas and pricing. Search indexes need refresh. prompts and tools change. Evaluation sets need maintenance. Secrets and identities expire or rotate. A candidate who can build a demo but cannot explain how the solution will be monitored, updated, secured, and supported still has an important readiness gap.

One final readiness check is whether you can explain tradeoffs without relying on the portal interface. Given a requirement, state why a model, retrieval approach, AI service, or agent pattern fits; identify the data and identity boundary; describe how quality will be measured; and name the failure you would test first. That explanation is a stronger signal of current engineering readiness than remembering the location of a setting.

Deployment should also be part of the lab rather than an afterthought. Put the application behind the identity and network controls you would actually use, record the dependencies it needs, and confirm what happens when one of those dependencies is unavailable. This exposes whether the design is operationally complete.

Keep model selection tied to a requirement. A larger or newer model is not automatically the better choice if a smaller model meets quality targets with lower latency and cost. Compare at least two viable options against the same evaluation set so model choice becomes an engineering decision rather than a preference.

AI-102 is useful history, but it is no longer the target. The current Azure AI engineering role is broader in generative AI, agents, retrieval, multimodal design, safety, and operations. If you carry forward the durable engineering knowledge and rebuild your labs around AI-103, the transition is manageable—and far more useful than continuing to prepare for an exam Microsoft has already retired.