Microsoft retired AI-102<\/a> and the Azure AI Engineer Associate certification on June 30, 2026. The direct replacement is AI-103, which earns Microsoft Certified: Azure AI Apps and Agents Developer Associate. Microsoft made the new certification available before the retirement so candidates could move to the new role without treating AI engineering as a discontinued track.<\/p>\n The short answer is therefore simple: AI-103 replaced AI-102. The useful answer is more nuanced. AI-103 is not a cosmetic renumbering. It shifts the center of the role toward Microsoft Foundry, generative AI, agents, retrieval, multimodal workloads, operational safeguards, and Python-based development while preserving several classic AI engineering responsibilities.<\/p>\n The replacement should be understood as a role transition rather than a simple exam-code substitution. AI-102 represented the Azure AI Engineer Associate path before Microsoft reorganized the role around Azure AI apps and agents. AI-103 keeps important engineering foundations but frames them around the current development platform and the workflows organizations are actually building. Someone who already passed AI-102 does not lose that engineering knowledge; someone preparing now, however, should not target the retired exam.<\/p>\n AI-102 should now be treated as historical context. Its final blueprint covered planning Azure AI solutions, generative AI, agentic solutions, computer vision, natural language processing, and knowledge mining or information extraction. That scope already reflected the move toward modern AI, but Microsoft replaced the credential rather than continuing to revise the old exam indefinitely.<\/p>\n AI-103 is the current role-based target for developers building Azure AI apps and agents. If a course still tells you to schedule AI-102, the course is stale. Older AI-102<\/a> material can still explain services and concepts that carry forward, but the current exam guide must be the source of truth for what to study now.<\/p>\n The new certification name matters. \u201cAzure AI Apps and Agents Developer Associate\u201d describes a builder who turns models into working systems. Candidates are expected to plan deployments, integrate data and tools, build generative and agentic experiences, monitor behavior, and apply security and responsible-AI controls.<\/p>\n That is different from thinking of an AI engineer primarily as someone who selects a cognitive service and connects an API. Modern solutions may use retrieval, vector search, multi-step reasoning, custom tools, conversational memory, multimodal inputs, and approval workflows. The application architecture around the model is now as important as the model call itself.<\/p>\n AI-103 is designed around Microsoft Foundry. Candidates need to choose appropriate models and services, configure projects and deployments, integrate retrieval and search, and connect the AI layer to application infrastructure. They also need to manage scaling, quotas, rate limits, cost, monitoring, and security.<\/p>\n This is one reason old AI-102 study plans can feel incomplete even when the concepts remain valid. The current role expects you to reason about the whole development surface: model selection, project configuration, SDK use, deployment topology, data grounding, safety controls, and production behavior. Passing the old exam was useful preparation, but it is not proof that those newer implementation patterns are familiar.<\/p>\n Foundry also changes the sequence in which many engineers learn Azure AI. Instead of studying one service family at a time, you are more likely to begin with a project, choose models and tools, connect knowledge, add an agent or application, and then operate the resulting system. The exam reflects that integrated workflow.<\/p>\n Cost and quotas deserve attention as engineering constraints. A solution can be functionally correct and still be unusable if token consumption, model throughput, search capacity, or concurrent agent activity cannot meet the workload. Capacity planning is therefore part of AI solution design rather than a finance exercise that happens after deployment.<\/p>\n That larger share changes how candidates should practice. It is no longer enough to know how to call a model endpoint. You need to reason about retrieval quality, grounding sources, conversation state, tool permissions, orchestration, evaluation, safety controls, latency, cost, and observability. Agentic systems also create security questions because a model can move from producing text to invoking actions. The engineering task becomes the design of a controlled system around the model, not just consumption of an AI service.<\/p>\n AI-103 gives substantial weight to generative AI and agentic solutions. Candidates should understand retrieval-augmented generation, tools, agent roles, memory, orchestration, evaluation, grounding, safety, and monitoring. Multi-agent and semiautonomous patterns are part of the current engineering conversation, not optional curiosities.<\/p>\n That does not mean every solution needs an agent. Good engineering still means choosing the simplest architecture that meets the requirement. The exam direction instead reflects the reality that developers must know when agentic patterns are appropriate, how to constrain them, and how to observe their behavior once they can call tools or act on enterprise data.<\/p>\n Tool access is especially important. An agent that can read a knowledge base is different from an agent that can modify a ticket, place an order, or change cloud resources. The second system needs explicit authorization, validation, approval, rollback, and audit design. Treating every tool as just another function call is a fast way to create an unsafe application.<\/p>\n The successor did not erase traditional Azure AI workloads. Computer vision, text analysis, speech, and information extraction remain part of the role. What changed is that these capabilities are increasingly composed with generative models and multimodal workflows rather than studied as completely separate product silos.<\/p>\n If you prepared for AI-102, this is one of the areas where your knowledge transfers well. The AI-102<\/a> experience around language, vision, responsible AI, and solution planning is still useful. Update the implementation details and learn how those capabilities are exposed and combined in the current Foundry ecosystem.<\/p>\n Practice Python in the context of Azure AI tasks rather than as a separate language course: parse configuration, call an SDK, handle API responses, transform retrieved data, evaluate output, and log failures. You do not need to become a language specialist before beginning, but you do need enough fluency to understand what the application is doing when an example stops working.<\/p>\n AI-103 expects development experience with Python. That expectation should shape your preparation. You need enough fluency to work with SDKs, build lightweight services, handle structured data, call APIs, manage configuration, and troubleshoot application behavior. Memorizing portal locations is not enough.<\/p>\n You do not need to become a language specialist. Focus on the patterns that appear repeatedly in AI application work: environment management, async calls, JSON, authentication, HTTP requests, SDK clients, error handling, logging, and testable functions. The stronger your basic software-engineering habits are, the easier it is to reason about agents and complex model workflows.<\/p>\nAI-102 is retired, and AI-103 is the current successor<\/h3>\n
The role is now explicitly about apps and agents<\/h3>\n
Microsoft Foundry is the organizing platform<\/h3>\n
Generative AI and agentic systems occupy a larger share of the job<\/h3>\n
Computer vision, text analysis, and information extraction still matter<\/h3>\n
Python is a real prerequisite, not a footnote<\/h3>\n
AI-300 is adjacent, not the replacement for AI-102<\/h3>\n