Most of your AI-102 prep still has value. The parts that don't are bigger than people are telling you.
Three of AI-102's six skill areas survive into AI-103 in name and largely in substance. One, the agentic sliver that was worth 5 to 10 percent, has been folded into the biggest domain on the new exam. And two areas kept their titles while having a large part of their content replaced underneath.
The single biggest cut: every custom-model-training objective is gone. Custom Vision training, LUIS-style language understanding, custom question answering, Video Indexer. None of it is examinable any more.
Below is the blueprint-level map, built from Microsoft's two published objective lists.
The short version
| AI-102 skill area | Weight | Becomes | New weight |
|---|---|---|---|
| Plan and manage an Azure AI solution | 20 to 25% | Plan and manage an Azure AI solution | 25 to 30% |
| Implement generative AI solutions | 15 to 20% | Merged into generative AI and agentic solutions | 30 to 35% combined |
| Implement an agentic solution | 5 to 10% | Merged into the above | see above |
| Implement computer vision solutions | 10 to 15% | Implement computer vision solutions, content largely replaced | 10 to 15% |
| Implement natural language processing solutions | 15 to 20% | Implement text analysis solutions, renamed and cut down | 10 to 15% |
| Implement knowledge mining and information extraction | 15 to 20% | Implement information extraction solutions, knowledge mining dropped from the title | 10 to 15% |
If you are still deciding whether AI-103 is the right exam at all rather than how to prepare for it, that is a different question, and the full objective breakdown lives in our AI-103 study guide.
Six areas became five. Sources: the AI-102 study guide, skills measured as of 23 December 2025, and the AI-103 study guide, skills measured as of 16 April 2026.
[IMAGE: Side-by-side screenshot of the two Microsoft Learn study guide pages, AI-102 and AI-103, both scrolled to the "Skills at a glance" list so the six areas and five areas are visible together. Alt text: Microsoft Learn study guides for AI-102 and AI-103 showing their skills measured lists side by side.]
What carries over unchanged?
Study these once, use them twice. If your notes cover the following, they are still current.
| Topic | AI-102 | AI-103 |
|---|---|---|
| RAG implementation | Implement a RAG pattern by grounding a model in your data | Implement retrieval-augmented generation in an application |
| Model selection and deployment | Choose the appropriate AI models, deploy using appropriate options | Choose an appropriate model for each task, configure model and agent deployments |
| Responsible AI and content safety | Content filters, blocklists, prompt shields, harm detection | Safety filters, guardrails, risk detection, content moderation |
| Azure AI Search indexing | Provision, create an index, define a skillset, semantic and vector | Ingest and index content, semantic, hybrid and vector search for grounding |
| Document extraction | Document Intelligence, Content Understanding | Content Understanding, OCR plus layout plus field extraction |
| Speech | Text to speech, speech to text, SSML, custom speech | Speech to text and text to speech for agentic interactions, custom speech models |
| Translation | Azure Translator, text and documents | Azure Translator in Foundry Tools or LLM-powered translation flows |
| Text analysis basics | Key phrases, entities, sentiment, PII detection | Entities, topics, summaries, sentiment, tone, sensitive content |
| Identity and cost | Managed identity, account keys, cost management | Managed identity, keyless credentials, quotas, rate limits, cost footprints |
| CI/CD | Integrate Foundry Services into a CI/CD pipeline | Integrate Foundry projects with CI/CD pipelines |
That is a substantial carry-over. Anyone claiming AI-103 is a completely different exam has not read both lists.
Two shifts worth noticing inside that table. Account keys became keyless credentials, which tells you which authentication pattern the exam now expects. And translation picked up "or LLM-powered translation flows", which means the service-first answer may no longer be the only right one.
Try a real Azure AI lab, free
A pre-provisioned Azure environment in your browser. No Azure account, no credit card, no setup.
Most large enterprises already run on Microsoft infrastructure. Azure AI Foundry is the platform those teams use to build and deploy AI apps, so the skills you practice here map directly to real job demand.
What is gone from AI-103 entirely?
This is where prep gets wasted. None of the following appears anywhere in the AI-103 objectives.
| Cut from the exam | Where it lived in AI-102 |
|---|---|
| Custom Vision model training: image classification versus object detection, labelling images, training, evaluating metrics, publishing, consuming | Implement custom vision models |
| Azure AI Video Indexer | Analyze videos |
| Spatial Analysis for detecting people in video | Analyze videos |
| Custom language models: intents, entities, utterances, training and deploying a language understanding model | Implement custom language models |
| Custom question answering: knowledge bases, question and answer pairs, multi-turn conversation, alternate phrasing, chit-chat, export | Implement custom language models |
| Custom translation training and publishing | Implement custom language models |
| Knowledge Store projections: file, object and table | Implement an Azure AI Search solution |
| Prompt flow | Build generative AI solutions with Microsoft Foundry |
| Container deployment for local and edge devices | Optimize and operationalize a generative AI solution |
| Fine-tuning a generative model | Optimize and operationalize a generative AI solution |
Look at the middle of that list. The entire "Implement custom language models" subsection of AI-102 has no successor. That was the LUIS lineage, and on AI-102 it sat inside a domain worth 15 to 20 percent. If your study plan had a week on intents, utterances and knowledge bases, that week is now unexaminable.
Same for Custom Vision. AI-103 still has a computer vision domain at the same weight, but it is about generating and understanding images rather than training a classifier.
One cut deserves separate mention because it did not disappear, it moved. Fine-tuning is no longer on AI-103. It is now on AI-300, the MLOps Engineer Associate exam, under "Implement advanced fine-tuning and model customization". If fine-tuning is central to your job, the exam that tests it is a different one now.
Tip
The fastest way to find out how much of this you already know is to try the work rather than read about it. The free taster lab is one Foundry deployment end to end, about twenty minutes, no card needed.
What is brand new with no AI-102 equivalent?
Everything here is an addition. If you sat AI-102, you have not seen any of it on an exam.
| New in AI-103 | Domain |
|---|---|
| Generating images from text prompts and reference media | Computer vision |
| Generating video from text prompts and reference media | Computer vision |
| Inpainting, mask-based edits, prompt-driven image modification | Computer vision |
| Editing generated video | Computer vision |
| Detecting indirect prompt injection via text embedded in images | Computer vision |
| Visual policy rules: watermarks, prohibited symbols, brand usage enforcement | Computer vision |
| Orchestrated multi-agent solutions | Generative AI and agentic |
| Agent roles, goals, conversation-tracking approach and tool schemas | Generative AI and agentic |
| Agent memory and function-calling integration | Generative AI and agentic |
| Autonomous and semi-autonomous workflows with approval flow controls | Generative AI and agentic |
| Governing agent behaviour with oversight modes, constraints and tool-access controls | Plan and manage |
| Model reflection, chain-of-thought evaluations, self-critique loops | Generative AI and agentic |
| Observability: tracing, token analytics, safety signals, latency breakdowns | Generative AI and agentic |
| Auditing through trace logging, provenance metadata and approval workflows | Plan and manage |
| Monitoring drift, safety events and grounding quality | Plan and manage |
| Monitoring data ingestion quality and search index health | Plan and manage |
| Private networking and role policies | Plan and manage |
| Single-task and pro-mode Content Understanding pipelines | Information extraction |
| Detecting fabrications when evaluating models and apps | Generative AI and agentic |
Count the agent rows. On AI-102, agents were one subsection called "Create custom agents" inside a domain worth 5 to 10 percent. On AI-103 they are woven through the largest domain on the paper and into the planning domain as well.
That is the real story of this transition, and it is not "AI-102 with newer service names".
Close the gap with hands-on Foundry labs
The Foundry, agent, and evaluation work that is new since AI-102, practised in real Azure environments.
So do your AI-102 study materials still work?
Partly, and it depends entirely on what kind of material you bought.
Official Microsoft learning paths. The AI-102 ones are retired. Microsoft's retirement announcement told candidates not close to testing to prepare for the new exam instead. Use the AI-103 paths.
Your own notes on services. Largely fine. Azure AI Search, Document Intelligence, Content Understanding, Speech and Translator all behave the same way and are still examinable.
Practice questions written for AI-102. Treat with suspicion. Any question on Custom Vision training, LUIS-style language understanding, question answering knowledge bases, Video Indexer or prompt flow is testing something that is no longer on the exam. A question bank that has not been rebuilt against the April 2026 objectives will spend your time on cut material.
Video courses recorded before 2026. Check the agent coverage. If the course treats agents as one short module at the end, it was built against the AI-102 blueprint and it under-weights the single largest thing AI-103 tests.
Anything mentioning "Azure AI Foundry". Not wrong, just dated. Microsoft's AI-102 change log records the objective wording changing from "Azure AI Foundry Services" to "Microsoft Foundry Services" on 23 December 2025. Both names refer to the same platform. Do not let a naming difference convince you a resource is out of date when it is otherwise accurate.
What should you actually re-study?
Three things, in this order.
One, agents as a system rather than a feature. Not "how do I create an agent" but how you define tool schemas, wire in retrieval and function calling, orchestrate several agents together, and put approval controls around an autonomous workflow. This is the largest addition and the one AI-102 barely touched.
Two, observability and evaluation. Tracing, token analytics, latency breakdowns, grounding quality, drift, detecting fabrications, provenance metadata for auditing. AI-102 asked you to enable tracing and collect feedback. AI-103 asks you to run a production system and prove it is behaving.
Three, the generative side of computer vision. If your vision knowledge is classification and object detection, it is pointed the wrong way. The new domain is generating images and video, editing them, and defending against prompt injection hidden inside them.
Notice what those three have in common. None of them is a fact you can memorise. They are all things you either have built or have not.
Go and open both study guides side by side, AI-102 and AI-103, and read the two computer vision domains against each other. Ten minutes on that comparison will tell you more about how much re-study you need than any summary, including this one.
Tip
When you have found your gaps, the AI-103 lab set covers the agent orchestration, evaluation and observability work objective by objective, on your own Azure subscription.