From AI-102 to AI-103: What Transfers and What You Need to Relearn

Microsoft AI-103 moves Azure AI development toward Microsoft Foundry, generative AI, retrieval, agents, evaluation, and multimodal applications. That center of gravity is different from many older Azure AI study plans, even though computer vision, language, speech, search, document extraction, security, and monitoring still matter.

The change is especially important for anyone coming from the retired AI-102, which retired on June 30, 2026. The active AI-103 blueprint, effective April 16, 2026, covers planning and managing Azure AI solutions, generative AI and agents, computer vision, text analysis, and information extraction, with Python experience and Microsoft Foundry familiarity expected. The current Azure AI Apps and Agents Developer Associate credential reflects that broader application-development role, while the retired Azure AI Engineer Associate is historical context rather than a current certification target.

What transfers, and what solution planning now requires

AI-102 candidates already had to choose Azure AI services, integrate them into applications, secure resources, monitor solutions, and think about responsible AI. Those habits still matter. AI-103 continues to ask which service or model fits a task, how an application should authenticate, how deployments should be monitored, and how safety and governance should be built into the solution.

Computer vision, language processing, speech, search, and document extraction have also not vanished. The current exam includes image and video generation and understanding, text analysis, translation, speech, retrieval, OCR, layout analysis, and structured information extraction. Candidates with genuine AI-102 hands-on experience therefore have a useful technical base.

The danger is assuming that because the service names look familiar, the new role is familiar too. AI-103 organizes more of the work around Foundry projects, models, retrieval, agents, evaluations, observability, and production lifecycle decisions. The old knowledge becomes an input to that system rather than the whole system.

Planning is a major current domain, accounting for 25–30 percent of the blueprint. The candidate must choose among large language models, small language models, multimodal models, Foundry Tools, retrieval approaches, indexing methods, memory options, tools, and knowledge sources. That is a broader design problem than selecting a single prebuilt AI service.

A useful planning sequence starts with the task. What input does the application receive? What output must it produce? Does the solution need grounding in private data? Is the task deterministic enough for a traditional service or rules engine, or does it need a generative model? Does the model need tools? Does the workflow require conversation memory? What latency, safety, privacy, and cost constraints apply?

Only after those questions should you select a model or Foundry capability. This prevents “model first” architecture, where a team picks a fashionable model and then tries to force every requirement through it. AI-103 rewards candidates who can match the component to the problem and then design the surrounding infrastructure, deployment, security, and monitoring.

For AI-102 to AI-103 readiness, mark old skills as transferable, changed, or new. Service integration may transfer strongly; Foundry project lifecycle, agent orchestration, and modern evaluation may require new practice.

Generative AI and agents are now core engineering work

The largest current domain, 30–35 percent, covers generative AI and agentic solutions. Candidates need to deploy and consume models, implement retrieval-augmented generation, design tool-augmented and multistep workflows, integrate Foundry SDKs and connectors, build agents with tools and memory, orchestrate multi-agent systems, and add safeguards and approval controls.

That changes how an AI-102 learner should practice. A prompt sent to a model endpoint is no longer enough. Build a small application that retrieves knowledge, passes grounded context to a model, calls at least one external tool, preserves appropriate conversation state, and records enough telemetry to understand what happened when the output is poor.

Understanding retrieval-augmented generation is especially important because the current blueprint connects retrieval, indexing, grounding, search quality, agents, and information extraction. A weak retrieval pipeline can produce a fluent but ungrounded answer even when the model itself is functioning normally.

Agents add another layer. You need to define roles, goals, tool schemas, conversation-tracking behavior, knowledge sources, and approval boundaries. Multi-agent systems add handoffs and orchestration. Once the basic distinction between a model call, a retrieval workflow, and an agent workflow is clear, study AI-103 generative AI and agents as one connected system rather than another list of features.

Evaluation, safety, and governance are part of development

Traditional software testing often asks whether an output is correct for a known input. Generative AI needs a broader evaluation model because quality can involve relevance, groundedness, fabrication, safety, latency, cost, and task success. AI-103 explicitly expects candidates to evaluate models and applications, monitor safety events and grounding quality, and perform error analysis on agents.

This means a production AI solution needs observable evidence. Trace model calls. Record retrieval behavior. Measure token use and latency. Track safety signals. Inspect failed tool calls and agent loops. Compare different prompts or model configurations against a repeatable evaluation set. The purpose is not to produce one attractive demo; it is to know why the application behaves the way it does.

AI evaluation becomes the feedback loop for architecture decisions. If retrieval relevance is weak, changing the prompt may not fix the problem. If the agent repeatedly chooses the wrong tool, the issue may be tool descriptions or orchestration. If a model is accurate but too slow or expensive, the right model choice may change.

Keep evaluation assets versioned alongside prompts, retrieval settings, and agent configuration. A small set of representative inputs, expected evidence, acceptance thresholds, and known failure cases makes regression visible when a model, index, prompt, or orchestration step changes.

AI-103 still includes responsible AI, but the current objectives go beyond generic principles. Candidates are expected to configure safety filters, guardrails, risk detection, content moderation, evaluators, trace logging, provenance metadata, approval workflows, oversight modes, constraints, and tool-access controls.

This matters because an agent can take actions, not just generate text. A poorly constrained tool call can be more damaging than an awkward answer. The solution therefore needs to control which tools the agent may invoke, what data it may access, when human approval is required, and how actions are audited.

Security is part of that architecture. The current planning domain includes managed identity, private networking, keyless credentials, and role policies. Avoid embedding credentials simply because a prototype is small. Treat model endpoints, search indexes, storage, tools, and agent identities as production resources with explicit access boundaries.

Vision, language, speech, and extraction still deserve real practice

The new exam is not only about generative AI. Computer vision, text analysis, and information extraction each account for 10–15 percent. The vision domain now includes image and video generation, multimodal understanding, captions, visual question answering, accessibility descriptions, Content Understanding, and visual safety controls.

Candidates who previously focused on classic image analysis should therefore update their practice. Current AI-103 computer vision work includes multimodal models and generated media as well as analysis.

Text analysis includes entity, topic, summary, sentiment, tone, safety, translation, domain outputs, speech-to-text, text-to-speech, custom speech, and audio reasoning. Information extraction combines ingestion, semantic/hybrid/vector search, enrichment, OCR, layout analysis, Content Understanding, and structured outputs for downstream reasoning.

These domains connect back to generative systems. Extracted documents can feed RAG. Visual understanding can supply evidence to an agent. Speech can become another interaction modality. Treat them as components in an application rather than isolated APIs.

Build one reference application that forces the old and new skills to meet

A practical transition project can cover more of AI-103 than a collection of disconnected demos. Start with a Python application connected to a Foundry project. Choose a model based on an explicit requirement. Add a small private knowledge set and build retrieval with semantic, hybrid, or vector search. Ground responses in the retrieved content and record citations or provenance.

Then add a tool. It can be a simple API or function that returns structured information. Define when the agent may call it and what happens if the call fails. Add conversation state only if the use case genuinely needs memory. Introduce a second agent or approval step if that improves the workflow rather than merely making the architecture look advanced.

Now instrument the application. Record traces, latency, token use, retrieval quality, tool calls, safety events, and failures. Create a small evaluation set and score relevance, groundedness, and task success. Add a visual, language, speech, or document-extraction component so the application touches at least one of the non-agent domains.

This project gives AI-102 veterans a way to preserve their Azure AI service knowledge while learning the new center of gravity. You are still integrating services and protecting resources, but now you are also designing retrieval, agents, evaluation, and operational controls as first-class parts of the solution.

The most important transition rule is therefore simple: do not study AI-103 as “AI-102 with new names.” Keep the durable engineering skills, retire the old exam assumptions, and rebuild your practice around Foundry, generative AI, agents, retrieval, evaluation, and multimodal applications. That is the difference between carrying useful experience forward and being trapped by a blueprint that no longer exists.