AI-103 vs AI-102: what changed?

Updated September 20, 2026

AI-103 replaced AI-102 when Microsoft retired AI-102 on June 30, 2026. If you studied for AI-102, a good part of your knowledge still applies, but the focus has shifted.

The main differences

AI-102AI-103
CertificationAzure AI Engineer AssociateAzure AI Apps and Agents Developer Associate
Centre of gravityIndividual Azure AI servicesGenerative AI and agents built with Microsoft Foundry
Largest domainSpread across servicesGenerative AI and agentic solutions (30–35%)
LanguageC# or PythonPython

What this means for your study plan

Spend more time on agents and RAG. You should be comfortable choosing the right tools, memory and knowledge sources for an agent, and grounding answers in your own data.

Learn to run AI in production. Planning and managing a solution carries 25–30% of the score. Expect questions about security (managed identity, private networking), monitoring, evaluation and cost.

Treat the classic services as consolidation. Vision, text analysis and information extraction are still on the exam, each at 10–15%. If you passed AI-102 you mostly need to refresh these and learn how they plug into agents.

Already certified for AI-102?

Your existing certification stays valid until its expiry date, but it can no longer be renewed on the old track. Most AI-102 holders move to AI-103 when their renewal comes up.

Which exam should you take now?

AI-103. AI-102 can no longer be booked, so the question is only relevant if you hold an AI-102 certificate or studied from older material. Anything written for AI-102, including books, videos and practice exams from before the switch, is a useful background but not a map of the current exam. Check the date on every resource and compare its outline with the AI-103 domains before you trust its weighting.

What carries over and what is new

AI-103 skill areaIf you know AI-102What to add
Plan and manage an Azure AI solution (25–30%)Resource creation, keys, endpoints and responsible AI basicsManaged identity, private networking, monitoring, evaluation and cost control for generative workloads
Generative AI and agentic solutions (30–35%)Calling a model deployment from codeAgents with tools and memory, grounding in your own data (RAG), orchestration with Microsoft Foundry
Computer vision (10–15%)Image analysis and OCR conceptsCheck the current service names and how vision plugs into an agent
Text analysis (10–15%)Language understanding, extraction and translation conceptsWhere a language model replaces a dedicated service, and where it does not
Information extraction (10–15%)Document and form extractionSearch and indexing patterns that feed a RAG solution

The pattern is simple. The classic services are the smaller part of the score, and you mostly refresh them. The new work is in the first two rows, which together carry more than half of the marks.

A three-week bridge plan

Week 1: the largest domain. Work through generative AI and agents explained, then build one small agent in Python that answers from a document set. A working lab teaches more than another day of reading.

Week 2: running it in production. Read plan and manage explained and apply it to the lab from week 1: give the app a managed identity instead of a key, add monitoring and write down what it costs per thousand requests.

Week 3: refresh and test. Skim computer vision, text analysis and information extraction, then sit the practice test under time pressure. The exam lasts 120 minutes, so practise at that pace. The full four-week study plan covers the same ground from scratch if you are not coming from AI-102.

Common mistakes when switching

  • Studying by service name. The exam asks you to pick a design for a scenario, not to recite what a service does. Practise choosing between options and explaining why.
  • Skipping the operations domain. Many developers treat monitoring, security and cost as someone else’s job. At 25–30% of the exam it is not.
  • Relying on old practice questions. Questions written for AI-102 overweight individual services and underweight agents. They will make you feel more ready than you are.
  • Assuming C# is enough. The new exam is Python-based, so be comfortable reading Python snippets even if C# is your main language.

Is AI-103 harder than AI-102?

It is broader in one direction and less spread out in another. Fewer separate services to memorise, but more design judgement about agents, grounding and production. If you already build with generative AI, it will feel closer to your job. If you mostly used the classic services, plan extra time for the first two domains. The is AI-103 hard page goes through the difficulty in detail, and is AI-103 worth it covers whether the certificate pays off.

For the wider picture of retired and replaced exams, see the replaced exams list and the exam changes tracker.