AI-103 Certification Guide: What Changed From AI-102

The AI-103 certification is Microsoft Certified: Azure AI Apps and Agents Developer Associate, and it is the exam to take now if you build AI applications on Azure. It replaced AI-102 (Azure AI Engineer Associate), which retired on June 30, 2026, and it moves the centre of the exam from calling individual Azure AI services to building generative AI apps and agents in Microsoft Foundry with Python.

This guide covers what the AI-103 exam is, what happened to AI-102 and its holders, how the skills outline changed, what each domain tests in practice, where AI-103 sits next to the other new Microsoft AI exams, and how to prepare without spending much. Every exam fact here was checked against Microsoft Learn on October 20, 2026; the announcement and replacement details come from Microsoft's Skills Hub blog.

AI-103 at a glance

The exam's full name is AI-103: Developing AI Apps and Agents on Azure. Passing it earns the Azure AI Apps and Agents Developer Associate certification, which Microsoft lists as intermediate level, for the AI engineer and developer roles. These are the facts Microsoft publishes, as of October 2026:

Status: live. Microsoft opened a beta in April 2026 and its June 2026 credentials roundup lists the exam as generally available. The certification page carries no beta label today.
Time: 120 minutes.
Questions: Microsoft doesn't publish a count for AI-103. Its general guidance is that most exams have 40 to 60 questions, and the page warns that you may meet interactive components.
Passing score: 700 on a scale of 1 to 1,000. It is a scaled score, so it does not mean 70% of the questions.
Prerequisites: none required. Microsoft expects experience building apps in Python and familiarity with general AI, generative AI and Azure services.
Languages: English, Chinese (Simplified and Traditional), French, German, Japanese, Korean, Italian, Portuguese (Brazil) and Spanish.
Delivery: proctored, scheduled through Pearson VUE.
Validity: one year, renewed for free with an online assessment on Microsoft Learn.

One detail surprises people. In role-based exams like this one, you can open Microsoft Learn during the exam, but the timer keeps running and you get no extra time. It helps with the odd lookup, not with learning a topic you skipped.

What happened to AI-102

Microsoft announced the change on its Skills Hub blog in March 2026, as part of a wider refresh of its AI certifications. AI-102, Designing and Implementing a Microsoft Azure AI Solution, retired on June 30, 2026, at 11:59 PM Central Standard Time. Its renewal assessment retired with it. Microsoft's retirement table names AI-103 as the replacement, and its advice to anyone who hadn't registered for AI-102 was to take the new exam instead.

If you already hold Azure AI Engineer Associate, three things are true according to Microsoft:

It stays valid until it expires. Retirement does not revoke an earned certification, and it stays on your Microsoft Learn transcript.
It can't be renewed any more. Renewal was possible only before the retirement date, so the credential lapses at the end of its current one-year term.
It doesn't turn into AI-103. Microsoft has published no conversion or upgrade path. To hold the current credential, you pass AI-103 like everyone else.

The same refresh replaced other exams, which is why AI-901, AI-200 and AI-300 appear in the same searches. They did not replace AI-102. More on them further down.

What changed from AI-102 to AI-103

The short version: agents moved from the edge of the exam to the middle, the per-service tasks shrank, and video generation and richer image editing arrived. Here is the detail, from the two study guides.

Domains and weights

AI-102's last outline, dated December 23, 2025, had six areas:

Plan and manage an Azure AI solution, 20-25%
Implement generative AI solutions, 15-20%
Implement an agentic solution, 5-10%
Implement computer vision solutions, 10-15%
Implement natural language processing solutions, 15-20%
Implement knowledge mining and information extraction solutions, 15-20%

AI-103's outline, current as of April 16, 2026, has five:

Plan and manage an Azure AI solution, 25-30%
Implement generative AI and agentic solutions, 30-35%
Implement computer vision solutions, 10-15%
Implement text analysis solutions, 10-15%
Implement information extraction solutions, 10-15%

Two moves matter. Generative AI and agents merged into one domain that is now the largest, where agents alone used to be worth 5-10%. And planning and management grew by five points; it holds the quotas, identity, networking and responsible AI controls that every production app and agent needs. Together those two domains are well over half the exam.

What's new or much bigger

AI-102 already touched agents, multi-agent orchestration, image generation with DALL-E, Content Understanding, tracing and CI/CD. What AI-103 adds is depth and a few genuinely new skills:

Agents in detail. Multi-agent orchestration was one line of a 5-10% domain; now the outline spells out agent roles, goals and tool schemas, conversation memory, semi-autonomous workflows with approval controls, and monitoring and error analysis for deployed agents.
Governance of agent behaviour. Oversight modes, constraints and tool-access controls, plus provenance metadata and approval workflows for auditing.
Video generation and richer image editing. Generating video from text prompts and reference media, editing generated video, and inpainting and mask-based image edits.
Multimodal understanding. Captions and alt text from multimodal models, question answering grounded in visual evidence, and video analysis with Content Understanding.
New risks. Detecting indirect prompt injection hidden as text inside images, and enforcing visual policy rules such as watermarks and brand usage.
Operations. Token analytics, safety signals and latency breakdowns in tracing, and chain-of-thought evaluations and self-critique loops.

What's gone or smaller

AI-103's outline no longer names training custom vision models, building custom language models with intents and entities, or question-answering knowledge bases, all of which were in AI-102. Azure Document Intelligence no longer has its own heading either; document extraction is described through Content Understanding and OCR pipelines. AI-102 asked for experience with Python or C#; AI-103's audience description names Python only.

Missing from the outline doesn't mean you will never see it. Microsoft notes that related topics may be covered, so treat the old services as background knowledge rather than study priorities.

What carries over

If you studied for AI-102, a good share of that work still counts: Azure AI Search with semantic, hybrid and vector search, enrichment skills and OCR, Azure Translator, speech to text and text to speech, sentiment and entity extraction, content filters, managed identity, private networking and quota management. They are now framed as parts of a generative or agentic solution rather than standalone services. Speech, for example, appears as "an agent modality".

The five domains in practice

Knowing the weights is one thing. Knowing what a question in each domain looks like is more useful, so here is what each domain asks you to decide, with a practice question from our bank for each.

Plan and manage an Azure AI solution (25-30%)

These questions give you a working app with an operational problem and ask for the right lever. The outline names deployment options, quotas, scaling and rate limits, keyless credentials and managed identity, private networking, and safety filters and guardrails. A typical one is a practice question on intermittent HTTP 429 errors during short traffic spikes, where usage averages well under the quota. The trap is reaching for more quota when the real issue is how the client handles bursts.

Implement generative AI and agentic solutions (30-35%)

This is the biggest domain and the one most changed from AI-102. You will choose between agent types, tools and patterns: a managed prompt agent versus a hosted one, an OpenAPI tool versus function calling, an MCP server versus a custom function, file search versus Azure AI Search. Try the practice question on choosing a prompt agent or a hosted agent for a team with no one to maintain code, and the practice question on calling an internal REST API from a managed agent. Human approval for risky tool calls is new territory, covered in a practice question on MCP tool approval. Evaluation sits here too: a practice question on picking a groundedness evaluator asks which metric catches answers that invent figures the retrieved chunks never contained.

Implement computer vision solutions (10-15%)

No more training custom vision models, and more generation and multimodal reasoning. Questions ask whether a task needs a generation model, an edit with a mask, a multimodal model that describes an image, or Content Understanding to pull structured fields from video. The practice question on mask-based image editing is a good example: change the background of thousands of product photos while leaving the product untouched.

Implement text analysis solutions (10-15%)

The outline calls this "language model text analysis". You extract entities, summaries and structured JSON with prompts and Foundry Tools, detect sentiment and sensitive content, translate with Azure Translator or an LLM flow, and wire speech in and out of agents. The usual choice is between a prompt to a model and a purpose-built tool, and the right answer depends on cost, consistency and whether a custom model or glossary is needed.

Implement information extraction solutions (10-15%)

This is the retrieval side of RAG. Questions cover indexing content in Azure AI Search, semantic, hybrid and vector search, enrichment skills, OCR and document extraction. The practice question on hybrid search with the semantic ranker shows a common pattern: pure vector search misses exact identifiers, so you add keyword matching and reranking. Another practice question on extracting fields from free-text letters covers the case where layouts vary and nothing is labelled. If RAG is new to you, our guides on what retrieval-augmented generation is and when to use RAG or fine-tuning cover the ideas the exam assumes.

Where AI-103 sits among Microsoft's new AI exams

Microsoft's 2026 refresh produced several exams with similar codes, and the Trends data shows people searching them together. Microsoft does not publish a required order, so here is what each one is, from its own pages:

AI-901, Azure AI Fundamentals. Replaced AI-900, which retired on June 30, 2026. A fundamentals exam with two domains, AI concepts (40-45%) and implementing AI solutions with Microsoft Foundry (55-60%). It is more hands-on than AI-900 was and expects basic Python. Fundamentals certifications don't expire. It is not a prerequisite for AI-103.
AI-103, Azure AI Apps and Agents Developer Associate. Building the AI app or agent itself.
AI-200, Azure AI Cloud Developer Associate. Replaced AZ-204. The back-end around AI apps: containers, Azure databases including vector search, messaging and eventing, security and monitoring. 120 minutes, 700 to pass.
AI-300, Machine Learning Operations Engineer Associate. Replaced DP-100. The operations side: MLOps and GenAIOps infrastructure, model lifecycle, quality assurance, observability and optimisation. 120 minutes, 700 to pass, English only. We have AI-300 practice tests too.

A reasonable reading: AI-901 if you are new to AI on Azure, AI-103 if you write the app or agent, AI-200 if you build the cloud services around it, and AI-300 if you deploy, evaluate and run models in production. That split is our reading of the outlines, not an order Microsoft sets.

Who should take AI-103

Developers adding AI features to products. The exam matches the work: retrieval, tools, agents, evaluation and safety, in Python on Foundry.
Azure AI Engineer Associate holders. Your credential runs out at the end of its current term with no renewal, so AI-103 is how you keep a current Microsoft AI associate cert on your profile. Expect to relearn the agent domain more than anything else.
People who passed AI-900 or AI-901. AI-103 is the natural next step if you code. If you don't, the fundamentals cert may be enough for your role.
Not a fit: if you never write code, or your work is operating ML and generative AI models in production. The first is better served by AI-901; the second by AI-300.

AI-103 exam cost and how to prepare

Microsoft prices exams by the country you take them in, and the AI-103 page doesn't print a figure; you see your local price when you schedule through Pearson VUE. Microsoft's FAQ says associate exams typically cost US$165, as of October 2026. Student pricing exists in most countries, though Microsoft's student discount page excludes India and China. An Exam Replay offer bundles the exam with one retake, both to be used within 12 months.

The free official resources are good and worth using in this order:

The AI-103 study guide on Microsoft Learn. The full skills outline. Read every bullet and mark the ones you couldn't explain to a colleague. That list is your study plan.
The training. The study guide points to free self-paced learning paths on Microsoft Learn, or an instructor-led course if you prefer a classroom.
Hands-on time in Foundry. Deploy a model, ground it with an index, wrap it in an agent with one tool, add an evaluator. Many questions assume you have seen those screens and SDK calls.
The free practice assessment. It is on AI Skills Navigator and needs a sign-in.
The exam sandbox. A demo of the real interface and item types, so the format holds no surprises.

Practice questions fill the gap between reading and the real exam. We have free AI-103 practice tests: ten tests of ten questions, each covering all five domains, with an explanation of why every wrong option is wrong and a link to the Microsoft Learn page behind it. You can also browse all 100 AI-103 practice questions by topic. When most tests pass comfortably, take the full-length practice exam on the AI-103 page: 50 questions in 120 minutes, answers hidden until you submit. These questions are original, written from Microsoft's public documentation; none is a real exam question.

Frequently asked questions

Is the AI-103 certification worth it?

If you build AI features on Azure, yes. It is the current Microsoft associate credential for that work, it tests agents and RAG rather than older single-service APIs, and renewal is free each year. It is less useful if your company doesn't use Azure, since the tools it tests are Microsoft's.

Is AI-103 harder than AI-102?

It is different more than harder. It is 20 minutes longer (120 against 100) and the pass mark is the same 700. The agent and generative AI material that was a minor part of AI-102 is now its largest domain, so people with AI-102 knowledge find the planning, search and language parts familiar and the agent questions new.

Can I still take AI-102?

No. AI-102 retired on June 30, 2026, and its renewal assessment retired with it. An Azure AI Engineer Associate you already earned stays valid until it expires.

How much does the AI-103 exam cost?

It depends on your country. Microsoft says associate exams typically cost US$165, as of October 2026, and shows the price for your country or region when you schedule.

Do I need AI-900 or AI-901 before AI-103?

No. Microsoft lists no required prerequisite. A fundamentals exam helps if you are new to AI, but it isn't required for the associate certification.

How long does it take to prepare for AI-103?

Microsoft doesn't say, and it depends on your starting point. A developer who already builds with Azure OpenAI or Foundry can usually focus on the agent and vision changes; someone new to Azure needs time for the planning domain as well. Taking one practice test before you start shows you which domains need the most time.