Home › Guides › AI-103 Exam Prep
Career Guide · 2026AI-103 Exam Prep 2026: A 6-Week Study Plan to Pass Azure's New Generative AI Certification
AI-103 exam prep 2026 means building a focused plan around Microsoft's replacement for AI-102: a 120-minute, 40 to 60 question test that needs a 700 out of 1000 to pass. A realistic six-week schedule splits time across five domains, weighted 10 to 35 percent, with the heaviest focus on building generative AI and agentic solutions on Microsoft Foundry.
- AI-102 stops being available on 30 June 2026, and it does not auto-upgrade to AI-103, so anyone still relying on an old study plan is preparing for the wrong exam.
- The pass mark is 700 out of 1000, scored on a scaled range, with 120 minutes to answer roughly 40 to 60 questions.
- One domain, generative AI and agentic solutions, carries 30 to 35 percent of the exam, more than double the weight of computer vision or text analysis alone.
- You do not need deep computer vision or NLP expertise to pass; those two domains combine for only about a quarter of the score.
- A working six-week plan spends three of those weeks on one domain (generative AI and agents) and treats the rest as a single combined review block.
- The exam costs 165 USD, payable in INR at the prevailing exchange rate through Pearson VUE, and retakes cost the same again.
- AI-103 pairs naturally with Claude's CCAR-F credential if your agents need to reason across both Azure and Claude tooling, which is increasingly how Indian GCCs are building.
Your manager just forwarded a Microsoft email: AI-102 retires on 30 June 2026, and if you want the replacement credential on your profile before your next appraisal cycle, you have about six weeks to get through five exam domains, one of which barely existed as a certification topic a year ago. That is the AI-103 exam, and it rewards a different kind of preparation than AI-102 ever did.
AI-103 Exam Prep 2026: What the Exam Actually Covers
AI-103, officially Exam AI-103: Developing AI Apps and Agents on Azure, is Microsoft's replacement for the retiring AI-102 (Azure AI Engineer Associate). It leads to the Azure AI Apps and Agents Developer Associate certification. The shift in name is not cosmetic. AI-102 tested whether you could wire together pre-built Azure Cognitive Services: Computer Vision, Language Understanding, Bot Service. AI-103 tests whether you can build a generative AI application or a multi-agent system on Microsoft Foundry (the platform formerly branded Azure AI Foundry), and whether you understand the guardrails around it.
Who the Azure AI-103 certification is actually for
If your day job involves calling an Azure OpenAI or Foundry model from an application, orchestrating more than one agent, or deciding how a RAG pipeline retrieves and grounds answers, AI-103 maps to your work. If your job is closer to classic .NET or Java development with the occasional API call, you will spend real study hours on unfamiliar territory: tool definitions, agent memory, and evaluation harnesses. A live, project-based generative AI developer course compresses that gap by making you build the exact kind of agent the exam asks about, rather than reading about it.
AI-103 at a glance
The numbers that should shape how you plan your six weeks.
Figures from Microsoft's published AI-103 exam guide, checked 21 September 2026.
AI-102 vs AI-103: What Actually Changed
If you already hold AI-102, it stays on your Microsoft transcript, but it will not renew after retirement and it does not convert to AI-103 automatically. You would sit AI-103 as a fresh exam, and honestly, the overlap is smaller than Microsoft's naming suggests.
| AI-102 (retiring 30 Jun 2026) | AI-103 | |
|---|---|---|
| Core platform tested | Individual Cognitive Services: Vision, Language, Bot Service | Microsoft Foundry, agent orchestration, generative AI apps |
| Heaviest domain | Implement computer vision solutions and NLP, spread fairly evenly | Generative AI and agentic solutions, 30 to 35% alone |
| New concepts tested | N/A | Agent memory, tool calling, multi-agent orchestration, safety evaluation |
| Exam format | ~120 min, ~40 to 60 questions | 120 min, 40 to 60 questions |
| Fee | $165 | $165 |
| Worth taking now? | Only if you can sit it before 30 Jun 2026 | Yes, this is the current path |
If you have three months of runway before that retirement date and your work is still mostly classic Cognitive Services, there is a case for squeezing in AI-102 first. For everyone else, and that is most people reading this in September 2026, AI-103 is the only sane target.
AI-103 Exam Domains and Weightings for 2026
Microsoft groups AI-103 into five domains. Two of them carry the exam; the other three are worth knowing well enough not to lose easy points, not worth over-investing in.
| Domain | Weight | What it actually tests |
|---|---|---|
| Implement generative AI and agentic solutions | 30 to 35% | Building apps and agents in Microsoft Foundry, RAG, SDKs and connectors, tool and memory design, multi-agent orchestration, safeguards, evaluation |
| Plan and manage an Azure AI solution | 25 to 30% | Project and deployment setup, CI/CD, quota and cost management, monitoring, managed identity, private networking, responsible AI controls |
| Computer vision | 10 to 15% | Image analysis, OCR, Vision SDK calls |
| Text analysis | 10 to 15% | Language service, sentiment, key phrase and entity extraction |
| Information extraction | 10 to 15% | Document Intelligence, structured extraction from unstructured files |
Add the top two domains and you are looking at 55 to 65% of the exam sitting on generative AI, agents, and solution operations. That is the honest reason a generic "Azure AI fundamentals" course will not get you through this: most of them were written to cover AI-102's discrete cognitive services, not agent design.
Is a Generative AI Developer Career Worth Building Around AI-103 in 2026?
Take Ananya, a 2021 computer science graduate working as a .NET developer at a mid-size insurer in Pune. Her team was told last quarter to bolt a chatbot onto the internal HR portal, then it quietly grew into "can it also answer policy questions from our PDF manuals." She has never touched Microsoft Foundry, never built a RAG pipeline, and has three weekends a month free.
That is a fair description of who AI-103 is built for right now: developers who got handed a generative AI feature before anyone formally trained them on one. Industry job listings suggest generative AI and agentic AI roles in India are among the fastest-growing postings on the major job boards this year, with pay for engineers who can show a working Foundry or agent project typically advertised above generic full-stack roles at the same experience level. Treat any specific figure you see quoted online as a rough band, not a guarantee: postings vary by city, company tier and whether the role is a genuine build role or an integration job with a fancy title.
Here is the honest caveat. AI-103 alone will not make you a generative AI developer. It certifies that you know Microsoft's tooling and vocabulary for the domain, which is useful for passing an HR keyword filter and for structuring your own learning, but it does not prove you can debug a RAG pipeline that returns confidently wrong answers, or that you can reason about why an agent looped three times on the same tool call. If your CV has the certification and nothing else, expect interviewers to probe past it within two questions.
One more distinction worth making before you commit six weeks to this: AI-103 is not the same target as AI-200. If your team's roadmap is mostly conversational AI features bolted onto existing apps rather than autonomous multi-step agents, the AI-200-focused Azure AI Cloud Developer course is the closer fit and a lighter lift. AI-103 earns its place when agents, not just chat features, are actually on your roadmap.
A 6-Week AI-103 Exam Prep Plan for Working Professionals
This assumes 6 to 8 hours a week, which is realistic for someone working full time. Ananya's version of this plan looked almost exactly like this, minus week 4, which she stretched to ten days because her day job had a release freeze.
The six-week block plan
Time allocated roughly in proportion to how each domain is weighted, not evenly.
Weeks 1 to 2: Foundry fundamentals
Stand up a Microsoft Foundry project, deploy a model, and wire a basic chat app through the SDK. Do not skip the console work for a tutorial video; the exam assumes you have clicked through project and deployment setup yourself.
Domain: Plan and manageWeeks 3 to 4: Generative AI and agents
Build one RAG pipeline end to end and one multi-step agent with at least two tools. Read the safety and evaluation documentation while you build, not after; the exam tests evaluation as part of the workflow, not as a separate topic.
Domain: 30 to 35% of the examWeek 5: The remaining three domains together
Computer vision, text analysis and information extraction combine for roughly a third of the exam. Do one small hands-on task in each, Vision SDK on a sample image, Language service on sample text, Document Intelligence on one invoice, rather than deep-diving any single one.
Domain: 30 to 45% combinedWeek 6: Practice exams and gap review
Take two full-length practice sets under timed conditions. Log every wrong answer by domain, not just by topic, so you can see whether your gap is concentrated in one weighted area before exam day.
Goal: identify weighted gapsA six to eight hour a week pace, built around the domain weightings above, checked 21 September 2026.
Skip the scattered tutorials and build the AI-103 projects with a mentor watching
360DT's Generative AI Developer course runs live over 8 weekends and maps directly onto the AI-103 domains: Foundry projects, RAG pipelines, and multi-agent builds, with prep for the Claude Certified Architect Foundations (CCAR-F) credential included. The next batch starts 27 September 2026.
Explore the course
Skills You Actually Need Before You Start
Do not open the exam guide first. Open your editor and check whether you can already do these five things; the gaps tell you where to spend week 1.
Microsoft Foundry certification skills you'll use on day one
Deploying a model, setting up a project and connection, and reading a quota error message without panicking. This sounds basic, but a surprising share of the "plan and manage" domain, worth 25 to 30% on its own, is really testing whether you have operated Foundry as an admin, not just called it as a developer.
Retrieval design, not just retrieval theory
You should be able to explain why a RAG answer came back wrong: was it a chunking problem, an embedding mismatch, or a prompt that buried the retrieved context under too much instruction. The AI Engineer course at 360DT spends real class time on exactly this kind of RAG debugging, which is a different skill from following a LangChain tutorial once.
Agent orchestration basics
Tool definitions, memory across turns, and what happens when an agent calls the wrong tool. If you have never built anything with more than one tool call, budget extra time here; it is the single most exam-heavy new topic versus AI-102.
Responsible AI, in Microsoft's specific vocabulary
Content filters, groundedness checks, and approval flows before an agent takes an action. This is a short list to memorise, and the exam rewards knowing Microsoft's exact terms for it over general AI safety knowledge.
Most people who fail AI-103 on the first attempt do not fail because they cannot build an agent. They fail because they treated the "plan and manage" domain as an afterthought, assuming it was the boring 25 to 30% they could wing. It is scenario-heavy: questions describe a deployment failure or a cost overrun and ask what you would configure, not what the feature is called. If you have never actually hit a quota limit or set up managed identity yourself, that domain will cost you more marks than the unfamiliar agent questions do.
The same gap shows up after you pass, not just on exam day. An agent that behaves in a demo can still drift, loop or hallucinate once it is handling real traffic, and monitoring that in production is a different skill from building the first version. If that is the direction your role is heading, the MLOps Engineer course covers the AI-300 material on exactly that operational layer, model and agent monitoring, retraining triggers, cost control, once AI-103 has you building.
Where the exam weight actually sits
The single biggest reason a leftover AI-102 study plan will not carry you through.
of AI-103 is generative AI and agentic solutions alone, a domain that did not exist on AI-102's blueprint.
Based on Microsoft's published AI-103 skills-measured outline, checked 21 September 2026.
AI-103 vs CCAR-F: Do You Need Both?
AI-103 proves you can build on Microsoft's stack. It says nothing about whether you can architect an agent responsibly on Claude, which is now the model many Indian GCCs and product teams are standardising on for reasoning-heavy agent work. The CCAR-F prep course covers the architecture side: how to design an agent's permissions, tool access, and escalation path regardless of which model sits behind it.
If your employer is Azure-first, AI-103 is the priority and CCAR-F is the upgrade you add once the first one is done. If you are job hunting rather than upskilling for a current employer, holding both signals something specific to a hiring manager: you are not locked into one vendor's agent framework, which matters more than either credential alone in a market where the tooling is still shifting every few months. Enterprises deploying agents also increasingly need someone who can govern them once they are live; if that administrative side interests you more than the build side, the Microsoft 365 Copilot Administrator course (AB-900) covers that adjacent track. The full certifications overview lays out how all of these paths connect if you are still deciding.
What a generative AI developer course covers that self-study skips
Self-study gets you through the multiple-choice questions. It rarely gets you through the "why did this agent do that" debugging conversation in a technical interview, because that requires having broken something in a live environment with a mentor who has seen the same failure before. That gap is exactly what a structured, live generative AI developer course is built to close, and it is why 360DT runs the AI-103 and CCAR-F material together rather than as two disconnected certificates.
A Realistic Six-Week Progression, If You Start Today
Here is roughly what Ananya's certification and career progression could look like if she starts this week rather than waiting for "a quieter month" that never arrives.
A six-week to four-month progression
Illustrative only: one plausible path from a standing start to a second, cross-platform credential.
Week 0
Week 4
Week 6
Month 4
An illustrative timeline, not a placement outcome or a guarantee, checked 21 September 2026.
If you are weighing whether to start, here is the call: do not wait for AI-102 to fully retire before you act, and do not try to self-study this one from documentation alone. The generative AI and agentic solutions domain is worth more than the other four combined, it is the newest material on the exam, and it is exactly the part a live, project-based course covers best. Treat the exam as the checkpoint, not the goal; the real asset is the working Foundry project and agent you build to pass it. If that is the six weeks you are about to spend, spend them inside 360DT's Generative AI Developer course, where the projects, the mentor feedback and the AI-103 domains are built around the same eight weekends.
Also read: Generative AI Developer Salary in India 2026 for how pay bands actually break down by experience and employer tier, and CCAR-F for Generative AI Developers if you want the architecture-side credential mapped out in detail.
Related guides
- How Do Large Language Models Work? the token and attention concepts underneath every Foundry model call the exam assumes you understand.
- What Is an AI Gateway in 2026? useful once your Foundry agent needs cost control and routing in production, not just in a demo.
- RAG vs Fine-Tuning in 2026 for the decision the "generative AI and agentic solutions" domain expects you to already have an opinion on.
- AI Engineer Roadmap 2026 if AI-103 turns out to be the first step in a broader move out of traditional development.
- What Is Context Engineering in 2026? the failure mode behind most of the agent bugs the exam's evaluation questions are really asking about.
Frequently asked questions
How do you prepare for the AI-103 exam in 2026?
Build hands-on projects in Microsoft Foundry across six weeks: two weeks on project and deployment fundamentals, two weeks on a RAG pipeline and a multi-tool agent, one week touching computer vision, text analysis and information extraction, and a final week of timed practice exams reviewed by domain, since the domains are weighted unevenly.
Is AI-103 harder than AI-102?
It tests newer, less standardised material. AI-102 covered mature, well-documented Cognitive Services; AI-103's biggest domain, generative AI and agentic solutions, covers a fast-moving area where fewer study resources exist yet, which makes it feel harder even though the exam format itself is unchanged.
When does AI-102 retire and can I still take it?
AI-102 is scheduled to retire on 30 June 2026. You can still schedule and sit it before that date, and any AI-102 certification you already hold stays on your transcript, but it will not renew after retirement and does not automatically convert to AI-103.
What is the passing score for AI-103?
Microsoft requires a scaled score of 700 out of 1000 to pass AI-103, in line with its other Associate-level certifications, across roughly 40 to 60 questions in a 120-minute session.
Do I need coding experience to pass AI-103?
Yes. AI-103 assumes you can work with an SDK to deploy models, call APIs, and wire up an application, even in scenario-based questions. It is not a conceptual, no-code exam; expect questions that describe code or configuration and ask what you would change.
How much does the AI-103 exam cost in India?
The exam fee is 165 USD, charged through Pearson VUE at the prevailing exchange rate on the day you pay, since Microsoft prices this exam in USD globally rather than in a separate INR band.
Is AI-103 or CCAR-F better for a generative AI developer career?
They are not competitors. AI-103 certifies Azure and Microsoft Foundry implementation skills; CCAR-F certifies agent architecture skills that apply regardless of the underlying model. Azure-first teams should prioritise AI-103 first and add CCAR-F once that is done.
What is Microsoft Foundry, and is it the same as Azure AI Foundry?
Microsoft Foundry is the current name for the platform previously branded Azure AI Foundry. It is Microsoft's unified environment for deploying models, building generative AI applications, and orchestrating agents, and it is the platform AI-103 is built around.
About this guide. 360 Digital Transformation is an Authorized Training Partner of Anthropic and Microsoft. Other certification bodies, vendors and employers named here are not affiliated with us. Figures cited were checked on 21 September 2026.
Microsoft Foundry
GPT and Foundry Models
Python
