Microsoft AI-103 After AI-102: What Azure AI Developers Need in 2026

Microsoft retired Azure AI Engineer Associate and exam AI-102 on June 30, 2026. The current successor path is Microsoft AI-103: Developing AI Apps and Agents on Azure, which supports the Azure AI Apps and Agents Developer Associate role.

The Microsoft AI-103 exam is now the correct primary target. The current blueprint measures planning and managing Azure AI solutions, generative AI and agentic solutions, computer vision, text analysis, and information extraction, with Microsoft Foundry at the center of the role.

The transition is not a simple rename. Some durable Azure AI skills still transfer, but AI-103 gives much more explicit weight to generative applications, agents, retrieval, evaluation, observability, safety, and production AI operations.

AI-102 is retired; AI-103 is the current developer path

Candidates should no longer build a new certification plan around retired AI-102 even though many Azure AI concepts from that exam remain useful. Within the Microsoft AI-102 to AI-103 transition, the ai-102 is retired; ai-103 is the current developer path decision matters because Microsoft changed both the credential and the role emphasis in 2026. A candidate working through ai-102 is retired; ai-103 is the current developer path should be able to state the governing requirement and identify the evidence that would prove the design, control, or transition choice is correct.

Audit old AI-102 notes and label each topic as directly transferable, changed in AI-103, or no longer central to the current blueprint. For a ai-102 is retired; ai-103 is the current developer path exercise, document the expected outcome and assumptions before acting. After the test, compare actual behavior with that expectation, then alter one meaningful condition and decide whether the original ai-102 is retired; ai-103 is the current developer path approach still fits.

Keeping a retired exam code as the visible study target can make otherwise useful technical material look current when the certification path is not. When reviewing ai-102 is retired; ai-103 is the current developer path, compare the preferred response with the nearest plausible alternative and name the exact condition that makes the alternative weaker. That ai-102 is retired; ai-103 is the current developer path boundary is more useful than memorizing a product, service, or exam phrase in isolation.

The AI-102 to AI-103 transition guide provides a detailed mapping.

Plan and manage Azure AI solutions — 25–30%

AI-103 expects developers to choose Foundry services, models, retrieval methods, deployment options, infrastructure, quotas, monitoring, security, and responsible AI controls. Within the Microsoft AI-102 to AI-103 transition, the plan and manage azure ai solutions 25 30% decision matters because production AI starts with architecture and operations rather than prompt code alone. A candidate working through plan and manage azure ai solutions 25 30% should be able to state the governing requirement and identify the evidence that would prove the design, control, or transition choice is correct.

Design one AI application and justify model type, Foundry project structure, deployment method, identity, networking, monitoring, cost, and safety controls. For a plan and manage azure ai solutions 25 30% exercise, document the expected outcome and assumptions before acting. After the test, compare actual behavior with that expectation, then alter one meaningful condition and decide whether the original plan and manage azure ai solutions 25 30% approach still fits.

Choosing a model before defining workload, grounding, security, and operational requirements can create a fragile solution. When reviewing plan and manage azure ai solutions 25 30%, compare the preferred response with the nearest plausible alternative and name the exact condition that makes the alternative weaker. That plan and manage azure ai solutions 25 30% boundary is more useful than memorizing a product, service, or exam phrase in isolation.

Generative AI and agents — 30–35%

The largest domain covers generative applications, retrieval-augmented generation, workflows, tools, agents, memory, multi-agent orchestration, evaluation, tracing, and operationalization. Within the Microsoft AI-102 to AI-103 transition, the generative ai and agents 30 35% decision matters because the current Azure AI developer role is centered on building and managing agentic and generative systems. A candidate working through generative ai and agents 30 35% should be able to state the governing requirement and identify the evidence that would prove the design, control, or transition choice is correct.

Build or design an agent that uses retrieval plus one external tool, then define conversation state, safeguards, approval, evaluation, and failure behavior. For a generative ai and agents 30 35% exercise, document the expected outcome and assumptions before acting. After the test, compare actual behavior with that expectation, then alter one meaningful condition and decide whether the original generative ai and agents 30 35% approach still fits.

A successful agent demo is not production-ready when tool permissions, grounding quality, observability, or approval controls are missing. When reviewing generative ai and agents 30 35%, compare the preferred response with the nearest plausible alternative and name the exact condition that makes the alternative weaker. That generative ai and agents 30 35% boundary is more useful than memorizing a product, service, or exam phrase in isolation.

The Microsoft AI-103 skills and architecture guide provides the broader current roadmap.

Computer vision — 10–15%

AI-103 includes image and video generation, multimodal understanding, visual question answering, accessibility outputs, content understanding, and responsible-AI controls for visual content. Within the Microsoft AI-102 to AI-103 transition, the computer vision 10 15% decision matters because modern AI applications increasingly combine text, image, video, and multimodal models. A candidate working through computer vision 10 15% should be able to state the governing requirement and identify the evidence that would prove the design, control, or transition choice is correct.

Design one multimodal workflow that accepts an image, extracts visual meaning, generates structured output, and applies safety checks. For a computer vision 10 15% exercise, document the expected outcome and assumptions before acting. After the test, compare actual behavior with that expectation, then alter one meaningful condition and decide whether the original computer vision 10 15% approach still fits.

Treating computer vision only as classification ignores the multimodal generation and understanding capabilities in the current role. When reviewing computer vision 10 15%, compare the preferred response with the nearest plausible alternative and name the exact condition that makes the alternative weaker. That computer vision 10 15% boundary is more useful than memorizing a product, service, or exam phrase in isolation.

Text analysis — 10–15%

The current role includes entity extraction, topics, summaries, structured outputs, sentiment, safety analysis, translation, speech, and domain-specific language workflows. Within the Microsoft AI-102 to AI-103 transition, the text analysis 10 15% decision matters because AI developers need to integrate language capabilities into broader applications and agents. A candidate working through text analysis 10 15% should be able to state the governing requirement and identify the evidence that would prove the design, control, or transition choice is correct.

Create a workflow that extracts structured information, summarizes it, evaluates sensitive content, and routes the result into a downstream application. For a text analysis 10 15% exercise, document the expected outcome and assumptions before acting. After the test, compare actual behavior with that expectation, then alter one meaningful condition and decide whether the original text analysis 10 15% approach still fits.

Using a general model without output constraints or evaluation can produce fluent but unreliable structured results. When reviewing text analysis 10 15%, compare the preferred response with the nearest plausible alternative and name the exact condition that makes the alternative weaker. That text analysis 10 15% boundary is more useful than memorizing a product, service, or exam phrase in isolation.

Information extraction — 10–15%

AI-103 covers ingestion, semantic and vector search, hybrid search, enrichment, OCR, RAG pipelines, Content Understanding, and structured extraction from documents and multimodal sources. Within the Microsoft AI-102 to AI-103 transition, the information extraction 10 15% decision matters because grounded agents and generative applications depend on reliable retrieval and clean source representations. A candidate working through information extraction 10 15% should be able to state the governing requirement and identify the evidence that would prove the design, control, or transition choice is correct.

Design an ingestion path from document to OCR or extraction, index, retrieval, grounded response, and evaluation of relevance. For a information extraction 10 15% exercise, document the expected outcome and assumptions before acting. After the test, compare actual behavior with that expectation, then alter one meaningful condition and decide whether the original information extraction 10 15% approach still fits.

A RAG system can fail because ingestion and retrieval are weak even when the language model is performing normally. When reviewing information extraction 10 15%, compare the preferred response with the nearest plausible alternative and name the exact condition that makes the alternative weaker. That information extraction 10 15% boundary is more useful than memorizing a product, service, or exam phrase in isolation.

Treat operations, security, and evaluation as cross-cutting concerns

Monitoring performance, safety events, grounding quality, search health, latency, token usage, identities, private networking, and audit evidence are integrated into AI-103 rather than optional production extras. Within the Microsoft AI-102 to AI-103 transition, the treat operations, security, and evaluation as cross-cutting concerns decision matters because AI systems can degrade or behave unsafely without traditional application errors. A candidate working through treat operations, security, and evaluation as cross-cutting concerns should be able to state the governing requirement and identify the evidence that would prove the design, control, or transition choice is correct.

Define metrics and alerts for one deployed agent, including quality, safety, cost, latency, tool use, retrieval, and failed approvals. For a treat operations, security, and evaluation as cross-cutting concerns exercise, document the expected outcome and assumptions before acting. After the test, compare actual behavior with that expectation, then alter one meaningful condition and decide whether the original treat operations, security, and evaluation as cross-cutting concerns approach still fits.

Monitoring only HTTP availability can miss an AI system that is online but producing low-quality or unsafe output. When reviewing treat operations, security, and evaluation as cross-cutting concerns, compare the preferred response with the nearest plausible alternative and name the exact condition that makes the alternative weaker. That treat operations, security, and evaluation as cross-cutting concerns boundary is more useful than memorizing a product, service, or exam phrase in isolation.

Use the transition to narrow, not expand, your study plan

Candidates with AI-102 experience should reuse durable Azure AI concepts while prioritizing the new AI-103 architecture, Foundry, agent, retrieval, evaluation, and operations emphasis. Within the Microsoft AI-102 to AI-103 transition, the use the transition to narrow, not expand, your study plan decision matters because starting from zero wastes prior knowledge, while keeping the old blueprint wastes time on outdated priorities. A candidate working through use the transition to narrow, not expand, your study plan should be able to state the governing requirement and identify the evidence that would prove the design, control, or transition choice is correct.

Create a three-column transition matrix: keep, update, and new, then attach one hands-on exercise to every high-priority gap. For a use the transition to narrow, not expand, your study plan exercise, document the expected outcome and assumptions before acting. After the test, compare actual behavior with that expectation, then alter one meaningful condition and decide whether the original use the transition to narrow, not expand, your study plan approach still fits.

Trying to master every Azure data and machine-learning service can distract from the current AI apps and agents role. When reviewing use the transition to narrow, not expand, your study plan, compare the preferred response with the nearest plausible alternative and name the exact condition that makes the alternative weaker. That use the transition to narrow, not expand, your study plan boundary is more useful than memorizing a product, service, or exam phrase in isolation.

The AI-102 to AI-103 practical transition exercises provide focused rehearsal.

The durable lesson from AI-102 is Azure AI engineering discipline; the current certification target is AI-103 and the role now centers on Foundry-based AI apps and agents.

Use the current five-domain AI-103 blueprint for final preparation and preserve old AI-102 material only where it supports those current responsibilities.