AI-103: Developing AI Apps and Agents on Azure
The replacement for the retired AI-102, and the first Microsoft exam with agents in its title — a third of the marks on generative and agentic solutions, and another 39% on the vision, language and extraction services most agent builders never touch.
- src
- Microsoft Learn — Azure AI Apps and Agents Developer Associate certification page (learn.microsoft.com/credentials/certifications/azure-ai-apps-and-agents-developer-associate)
- chk
AI-102 retired on 30 June 2026 and this is what replaced it. The rename is the summary: Azure AI Engineer became Azure AI Apps and Agents Developer, and the exam moved with it.
Where the marks actually are
The largest domain is Implement generative AI and agentic solutions at 30–35%. Planning and managing the solution is 25–30%, and that one is less glamorous than it sounds — quotas, rate limits, cost footprints, managed identity, private networking, and the governance controls that decide what an agent is allowed to call.
Then three bands of 10–15% each: computer vision, text analysis, information extraction.
Add those three up and 39% of the exam is the older Azure AI services surface — OCR, sentiment, speech, image generation, document extraction. That is the trap. People book this exam because it has agents in the title, prepare the agentic third thoroughly, and discover that two of every five marks are on services they have never opened.
Microsoft publishes ranges, not weights
The official objectives give bands — 25–30%, 30–35% — rather than fixed percentages. The figures on this page are the midpoints normalised to 100, with every value kept inside its official range. Microsoft's study guide is the authoritative source, and the ordering does not change either way.
This is the same treatment AZ-104 gets on this site, and it is worth stating rather than hiding: a page that silently invents exact numbers from a band is asserting precision the vendor did not publish.
Python is a real prerequisite
There is no prerequisite exam, and there is a stated language requirement: experience developing apps using Python. The agentic domain asks for tool schemas, function calling, Foundry SDK integration and multi-agent orchestration. None of that is portal work.
This is the clearest difference from the AZ-104 and SC-500 side of Microsoft's catalogue, where a competent administrator can pass without writing code. Here, an administrator cannot.
The renewal cycle is the hidden cost
Twelve months, against two years for the Linux Foundation and CNCF exams. Renewal is free and unproctored — an online assessment on Microsoft Learn from six months before expiry — but it is an annual obligation, and in a field moving as fast as this one the renewal assessment will not be testing what you sat.
Treat that as a feature rather than a tax. An annually refreshed credential in agentic AI says something a two-year-old badge cannot.
Before you book
Build one agent that calls a tool and asks a human first. Approval flows and oversight modes are named competencies in two separate domains, and they are the part that separates a demo from something an employer would deploy.
Do not skip vision and speech. Thirty-nine per cent. A perfect agentic score with nothing here does not reach 700.
Check the objectives date on your booking. The current set is dated 16 April 2026, and Microsoft runs two versions in parallel whenever an exam changes.
Exam domains
Implement generative AI and agentic solutions
33%Plan and manage an Azure AI solution
28%Implement computer vision solutions
13%Implement text analysis solutions
13%Implement information extraction solutions
13%Preparation path
- 1
Read the study guide and check which version of the exam you are sitting
Microsoft publishes two sets of objectives when an exam changes, and AI-103 is young enough that the distinction matters. The current set is dated 16 April 2026. Confirm which applies to your booking before you plan anything else.
~3 hours - 2
Build an agent in Microsoft Foundry with tools and memory — this is the largest domain
A third of the marks. Define the roles and tool schemas, wire function calling and conversation memory, then orchestrate two agents rather than one. The blueprint asks for approval flows and safeguards explicitly, so build the human-in-the-loop step too.
~30 hours - 3
Ground a model with Azure AI Search, then prove the grounding works
RAG appears in the generative domain and again in information extraction, so the retrieval pipeline is worth more than its 13% suggests. Build semantic, hybrid and vector search over the same corpus and compare what each returns for the same question.
~20 hours - 4
Plan the boring half — quotas, identity, private networking and cost
The second-heaviest domain at 28% is mostly operations, not model work: quotas, rate limits, cost footprints, managed identity, keyless credentials and private networking. It is the domain a working AI developer is most likely to have skipped entirely.
~16 hours - 5
Wire safety filters and evaluators, including the prompt-injection cases
Responsible AI is not a separate domain — it is threaded through planning, vision and text. Content Safety, evaluators and the indirect prompt injection hidden in an image are all named competencies, and the last one surprises people.
~12 hours - 6
Spend a weekend on vision, language and speech even if you never use them
Thirty-nine per cent of the exam is computer vision, text analysis and information extraction. Most people building agents touch none of it, and skipping it caps you below the 700 mark even with a perfect agentic score. Breadth beats depth here.
~18 hours
Frequently asked questions
Career Roadmaps
- AI Agents Engineer RoadmapA path from LLM and programming fundamentals through agent orchestration, tool use, and production deployment for building autonomous AI agent systems.
- AI Security Engineer RoadmapA defensive security path for engineers who secure LLM and agent systems, covering AI threat modelling, prompt injection defence, supply chain integrity, agent permissions, guardrails, governance and incident response.
- Cloud Architect RoadmapA path into cloud architecture as the job it actually is — trade-off analysis, migration of systems you did not write, disaster recovery you have rehearsed, decision records, and influence without formal authority.