Home › Certifications › AI-103 for Generative AI Developers
Microsoft Certification Guide · 2026AI-103 for Generative AI Developers: The 2026 Guide
The half of AI-103 that agent-focused guides skip. Why multimodal and information extraction alone are up to 45% of the exam before you count grounding, how RAG is actually assessed, why responsible AI is threaded through every domain rather than isolated, what a generative AI developer builds that an agent engineer does not, and a 12-week roadmap.
Training & Technology Partners
Microsoft
AI-103 — Azure AI Apps & Agents Developer Associate
The credential this guide covers
Claude
Claude Certified Architect – Foundations (CCAR-F)
The second credential in this programme — read the CCAR-F prep guideEveryone writes about the agentic domain on AI-103, because it is the biggest single block at 30–35%. But a generative AI developer spends most of their working life somewhere else — grounding models in real data, handling documents, images and speech, and proving the output is safe. Those live in the three domains at 10–15% each, plus the retrieval half of the generative domain. Together that is comfortably half the exam, and it is the half that gets under-studied.
This is the third angle we publish on AI-103, and deliberately the least agent-centric. Our complete AI-103 certification guide walks the whole blueprint; the agentic engineering guide reads the exam through orchestration and tool calling. This one is for the developer whose job is shipping generative AI features into a product — a grounded assistant, a document pipeline, a multimodal search — where the agent, if there is one at all, is a small part of a much larger system.
Two different jobs behind one exam
AI-103 certifies Azure AI Apps and Agents Developer — and those really are two jobs. Recognising which one you are hiring for, or being hired for, changes how you read the blueprint.
| Generative AI developer | Agentic AI engineer | |
|---|---|---|
| Ships | Features inside a product — assistants, summarisers, document pipelines, multimodal search | Systems that plan and act — bounded loops, tool suites, multi-agent orchestration |
| Hardest problem | Retrieval quality — getting the right context in front of the model | Loop control — stopping the agent doing the wrong thing repeatedly |
| Fails by | Confident, fluent, wrong answers grounded in the wrong chunk | Infinite retries, tool misselection, context drift over long runs |
| AI-103 domains | Vision, text analysis, extraction, plus the grounding and RAG half of the generative domain | The agentic half of the generative domain, plus tool integration |
| Under-studies | Orchestration patterns and tool schemas | Vision, text analysis and extraction — up to 45% of the exam |
AI-103 domains, read from the generative side
The same five domains, ordered by how much of a generative AI developer's day they describe.
AI-103 exam blueprint — share of content by domain
Official Microsoft weightings, published as ranges
Source: Microsoft Learn, AI-103 skills measured. The three "small" domains total 30–45% between them — potentially more than the headline generative domain. Bar length maps each range's midpoint against the largest domain's midpoint. Because these are ranges, midpoints sum to roughly 100% but any sitting can shift within the bands.
If you study only the agentic domain you have covered at most 35%. Add platform management and you are at 65%, still short of a comfortable pass. Vision, text analysis and information extraction are up to 45% combined — and unlike orchestration, they are largely recall: which service, which feature, which input format. They are the cheapest marks on the exam and the ones most often left on the table.
Grounding is the job, not a sub-topic
For a generative AI developer, retrieval quality determines product quality. A model with excellent reasoning and the wrong three paragraphs of context produces a confident, fluent, wrong answer — and users trust it, which is worse than an obvious failure.
Chunking beats embedding choice
Teams agonise over embedding models and accept whatever chunking their framework defaults to. That is backwards. Chunk size, overlap and whether you split on structure or on token count change retrieval quality far more than swapping one competent embedding model for another. Get chunking right first.
Grounding & RAGMeasure retrieval separately
When a grounded assistant answers badly, the cause is usually retrieval, not generation — but teams debug the prompt. Evaluate retrieval on its own: for a set of real questions, did the right passage come back at all, and at what rank? Fix that number before touching the prompt.
EvaluationHybrid search and reranking
Pure vector search misses exact matches — product codes, names, error strings — that keyword search finds instantly. Hybrid retrieval plus a reranking pass is the standard production answer, and knowing why each half exists is more useful on the exam than memorising either.
Azure AI SearchHandle the empty result honestly
What does your system do when nothing relevant is found? The wrong answer is to let the model improvise. Detecting low-confidence retrieval and saying so is a design decision, and it separates a product people trust from a demo that impresses once.
ReliabilityCitations are an architecture choice
Traceability back to the source is not a feature you bolt on later — it constrains how you chunk, what metadata you store, and what the model is asked to return. Decide it at design time. It is also what makes a grounded system auditable, which is increasingly a procurement requirement.
TraceabilityMultimodal is not a bonus round
Real document pipelines are not plain text. Scanned PDFs, tables, forms, images, recorded calls. Vision, speech and translation plus information extraction are where enterprise generative AI actually lives — and they are three separate examined domains. Treat them as core, not as trivia.
Up to 45% combinedResponsible AI is threaded, not bolted on
Microsoft does not isolate responsible AI into its own domain — it appears inside otherwise-technical questions across the blueprint. For a generative AI developer that mirrors reality: content filtering, harm categories, groundedness checks and monitoring are part of shipping the feature, not a compliance review afterwards. Expect questions where the technically capable option is wrong because it lacks a safety control the scenario requires.
AI-103 exam at a glance
| Attribute | Detail |
|---|---|
| Exam code | AI-103 — Developing AI Apps and Agents on Azure |
| Certification earned | Microsoft Certified: Azure AI Apps and Agents Developer Associate |
| Level | Intermediate / Associate |
| Duration | 120 minutes, proctored, may include interactive lab-style components |
| Passing score | 700 out of 1000 |
| Delivery | Pearson VUE — online proctored or at a test centre |
| Prerequisites | None enforced — but Python experience and Azure familiarity are assumed |
| Renewal | Expires annually; renew free via an online assessment on Microsoft Learn |
| Replaces | AI-102 (Azure AI Engineer Associate), retired 30 June 2026 |
| Paired with | Claude Certified Architect – Foundations (CCAR-F) in the 360DT Generative AI Developer programme |
Who this angle is for
- Developers shipping AI features into an existing product — where the AI is one component of a real application, not the whole thing.
- AI-102 holders whose credential retired on 30 June 2026 and who need the current equivalent.
- Backend and data engineers — RAG pipelines, document processing and vector search are closer to data engineering than to machine learning.
- Anyone who wants dual Microsoft and Anthropic credentials — the platform certification plus the vendor-neutral architecture one, which is what our programme runs.
Where it will not carry you
- It does not make you an ML engineer. You consume models rather than train them — that path is AI-300.
- It is Azure-specific. RAG and evaluation concepts travel; Microsoft Foundry and Azure AI Search do not.
- It will not substitute for writing code. No prerequisite is enforced, but the exam assumes Python and hands-on Azure work.
What a generative AI developer earns
Generative AI development sits where ordinary software engineering meets a scarce specialism. The premium goes to people who can ship a grounded system that holds up in production — not to people who can call an API.
The AI developer progression
Entry
Mid
Senior
Lead / Architect
AI developer pay by market
India and the US shown as two separate charts, because rupee and dollar bands are different measures on different scales and should never share an axis.
India — annual CTC
AI and generative AI developer roles, ₹ lakh per annum
Bar length maps the upper bound of each band against a ₹53 L scale. Bands compiled from published Indian AI engineer ranges. Product companies and global capability centres sit at the top of each band; service firms at the bottom.
United States — annual base
AI developer roles, US$ thousands
Bar length maps each figure against a $160K scale. Bands compiled from published US AI engineer ranges; base pay only, excluding bonus and equity.
These are role bands, not certification outcomes. No published figure attaches a salary increase to holding AI-103 alone. The 45% figure above is the top of Microsoft's published ranges for the three domains combined and describes exam weighting, not earnings. What is defensible: a certification plus shipped, grounded systems you can demonstrate is a far stronger application than either on its own.
Compensation figures are compiled from independent, publicly available industry sources and are shown for role context. They are not a guarantee of pay in any specific market, company or outcome, and 360DT does not promise a salary result from any certification or programme.
Your 12-week roadmap
Front-loaded on the platform and grounding, with the three "small" domains given a proper block rather than a weekend. Assume 8–10 hours per week and an Azure account open throughout.
Foundry and the platform floor
Microsoft Foundry: projects, deployments, model selection. Authentication and keys. Which Azure AI service does what, and where responsibilities overlap. Unglamorous, and 25–30% of the exam.
Cost, quota, monitoring and deployment
The rest of the platform domain: provisioning, scaling, quota management, and monitoring a solution you have shipped. This is the half of "plan and manage" that developers assume is someone else's job. On this exam it is yours.
Generative development in Foundry
Prompt patterns that survive being called a thousand times. Structured output and why everything downstream depends on it. Model selection against cost and latency, not just capability.
Retrieval: chunking and vector search
Chunk size, overlap, and splitting on structure versus token count. Embeddings and vector stores. Build a retrieval pipeline over a real corpus — ideally your employer's documents, because sanitised sample data hides every interesting problem.
Grounding on Azure AI Search
Hybrid search, reranking, filters and metadata. Citations and traceability. Then measure retrieval on its own — right passage returned, at what rank — before you touch a prompt. This week is the core of the job.
Agents, lightly
Memory, tool calling and enough multi-agent orchestration to answer the questions confidently. You do not need agentic depth for this angle — but it is part of the largest domain, so cover it properly once. Our agentic guide goes deeper if you want it.
Safety, evaluation and observability
Content filtering and harm categories. Groundedness checks. Evaluation harnesses for something non-deterministic. Logging and tracing a generative feature in production. Responsible AI runs through every domain, so this pays back everywhere.
Computer vision
Image analysis, OCR on real scanned documents, and the input formats and limits that trip people up. One of three domains worth up to 15% each, and largely recall — cheap marks if you sit down and learn them.
Text analysis, speech and translation
Sentiment, entities, key phrases, language detection. Speech to text and translation for multilingual products. Where each fits against a generative model — because the exam expects you to know when a purpose-built service beats a prompt.
Information extraction
Structured data out of unstructured documents: forms, tables, invoices, contracts. Custom models and when they beat a general one. Completes the three domains that are up to 45% between them.
Build one grounded system end to end
Real corpus, hybrid retrieval with reranking, citations, structured output, a safety filter and an evaluation harness. Handle the empty-result case honestly. This is simultaneously exam revision and the portfolio piece an interviewer will actually ask about.
Timed practice, then the exam
Full 120-minute sittings scored per domain, not overall. Anything under 70% in a domain gets a dedicated re-study day. Rehearse the interactive lab-style components separately — they eat more time than you expect. Then book.
Read every question for the constraint — latency, cost, data residency, least operational overhead. Two options will usually both work and only one satisfies it. And watch for questions where the technically capable answer is wrong because it omits a safety control the scenario requires: responsible AI is threaded through every domain rather than isolated in its own.
Learn from Microsoft and Anthropic certified trainers
Certification prep at 360DT is delivered live by AWS, Microsoft, Cisco and NVIDIA certified trainers and Instructor at 360DT, an Anthropic Partners — our full international faculty.
Naga Nandhini
Java Full Stack Corporate Trainer and Enterprise Solutions Expert
Powerhouse corporate trainer with 10+ years of upskilling engineers at HCL, Accenture, Cybage, Mobileum and Renault. An award-winning engineer who helped build the RAPTOR and KIKO platforms for an American Airlines client, she brings real enterprise delivery into every session.
Praveen Kumar
Full Stack and GenAI Corporate Trainer, Ex-TCS Team Lead
Full stack lead and GenAI builder with 10+ years of training behind 50+ corporate batches for TCS, HCL, Perficient, Renault Nissan and Saint-Gobain. He ships real RAG and local-LLM systems with Ollama, FastAPI and Spring Boot, then teaches exactly how they work.
Bipeen
AWS, Microsoft, and Anthropic Certified Trainer
AI, machine learning, and cloud transformation expert with more than 25 years of experience. A renowned global corporate trainer and conference speaker known for practical, outcome driven sessions.
Florian Garcia Compte
Cisco Certified Systems Instructor (CCSI)
Partner Director at AVAANZA FORMACION, a Cisco Learning Partner based in Madrid. Certified Cisco Systems Instructor (CCSI No. 21053) specializing in Cisco Data Center (ACI/SDN), Nexus, and Wireless, delivering official Cisco certification courses including CCNA, CCNP Enterprise, and CCNP Data Center.
Shantanu Pandey
Microsoft, Google, NVIDIA and HPE Certified Trainer
Director of Engineering at Meteoros Automation, ranked in Microsoft's Top 100 Trainers of 2025, holding MCT, Google Cloud, NVIDIA and PeopleCert credentials, with 500+ trainings delivered since 2011.
Vikas Mittal
AI instructor with 25+ years in technology training
Veteran technology trainer with more than 25 years of rich industry experience spanning Google, Microsoft, and emerging AI platforms. An exceptional mentor recognized for proven teaching excellence.
Sid
Microsoft Certified Trainer since 2016 and Azure Data Engineer
Dynamic corporate trainer and consultant with more than 10 years of experience across Microsoft Fabric, Azure Data Engineering, and modern cloud platforms. Focused on clear, practical learning that sticks.
Arshad Ahmad
Microsoft Certified Trainer (MCT), Cybersecurity & Power Platform Expert
Technology Trainer with 15 years in strategic client delivery, holding MCT, Azure Solutions Architect Expert, Cybersecurity Architect Expert and Power BI Data Analyst credentials, with 200+ trainings delivered.
Sravia
NVIDIA Authorized Instructor and AI Data Specialist
Passionate AI and data evangelist with over 10 years of experience across machine learning, MLOps, and cloud native technologies. Dedicated to delivering impactful, engaging enterprise training programs.
Johan
Microsoft Business Applications MVP and Microsoft Certified Trainer
D365 Customer Service Lead – Europe at Avanade, and founder of the Power Platform School. A Microsoft Business Applications MVP and Microsoft Certified Trainer based in London, specializing in Dynamics 365 Customer Service and Power Platform.
Akim Nyamande
Microsoft, CompTIA, Juniper and Cisco Certified Trainer
IT training facilitator with over five years of experience and certifications across Microsoft, CompTIA, Juniper, and Cisco. Previously a Network Administrator before moving into technical training, now delivering hands-on courses in networking, systems administration, and cybersecurity fundamentals.
Ali El Khatib
Cisco Certified Systems Instructor (CCSI) and Microsoft Certified Trainer
Infrastructure Engineer at RHUH with 13 years in the training field, holding CCSI, MCSE, CCNP Routing and Switching, CCNP Security, CompTIA and Microsoft Azure certifications.
Generative AI Developer Program
The 360DT Generative AI Developer Program runs the full AI-103 blueprint — Foundry, grounding and RAG on Azure AI Search, agents and orchestration, safety and evaluation, multimodal, and information extraction — then carries on into Claude Certified Architect – Foundations (CCAR-F). You finish with five shipped systems, including a grounded enterprise assistant, a chunking and retrieval benchmark, and an evaluation and safety harness.
Explore the Generative AI Developer Program
Generative AI developer and AI-103 frequently asked questions
Is AI-103 worth it if I am not building agents?
Yes — arguably more so. Despite the name, only 30–35% of the exam is the combined generative and agentic domain, and much of that is grounding rather than orchestration. Platform management is 25–30%, and vision, text analysis and information extraction are up to 45% between them. The majority of AI-103 describes generative application development, not agent engineering.
How is this different from your other AI-103 guides?
Three angles on the same exam. Our complete AI-103 guide is the full walkthrough of every domain. The agentic engineering guide reads it through orchestration, tool calling and MCP. This one is for developers shipping generative features into products — grounding, multimodal and extraction.
What matters more, the embedding model or the chunking?
Chunking, in most real systems. Chunk size, overlap and whether you split on document structure or raw token count affect retrieval quality more than swapping between competent embedding models. Teams routinely spend weeks on model selection while accepting their framework's default chunking, which is the wrong order.
Do I need Azure AI Search, or can I use any vector database?
In production, any competent vector store works and the concepts transfer. For the exam, Azure AI Search specifically — AI-103 examines Microsoft services, so hybrid search, semantic ranking, filters and index design as Azure implements them are what you will be asked about.
Which certification does this programme pair AI-103 with?
Claude Certified Architect – Foundations (CCAR-F), Anthropic's architecture-tier credential. Some older listings abbreviate it CCA-F or CCFA; CCAR-F is the current Pearson VUE exam code. Our CCAR-F prep guide covers how to pass it.
I hold AI-102 — is it still valid?
AI-102, the Azure AI Engineer Associate exam, retired on 30 June 2026. AI-103 is its replacement and reflects a materially different blueprint, with agents named in the certification title and generative development given far more weight. If your AI-102 credential is still inside its annual renewal window it remains valid, but new candidates take AI-103.
How long does preparation take?
About twelve weeks at 8–10 hours per week on the roadmap above. Developers already working on Azure daily often compress it to eight. The programme runs 8 weeks live with 60+ hours of mentor-led instruction and covers both credentials.
Are the exams included in the programme fee?
No. Certification exams are administered independently by Microsoft and Anthropic and are not included in programme tuition. The fee of ₹24,999 covers the live training, the five projects and certification preparation; you book and pay for each exam directly with its issuer.
Sources and further reading
- Microsoft Learn — Exam AI-103: Developing AI Apps and Agents on Azure, skills measured (five domains and their published weighting ranges, 120-minute duration, 700 passing score, interactive components, annual renewal)
- Microsoft Learn — Azure AI Apps and Agents Developer Associate certification page, and the AI-102 retirement notice (30 June 2026)
- Anthropic Partner Academy — Claude Certified Architect – Foundations, the second credential in this programme
- Pearson VUE — Claude Certification Program by Anthropic, used to confirm CCAR-F as the current exam code
- The combined 30–45% figure for vision, text analysis and information extraction is the sum of Microsoft's three published ranges, calculated by us
- Published Indian and US AI engineer and generative AI developer compensation ranges, used for role context only
360 Digital Transformation is an Anthropic Partner and a Microsoft Authorized Training Partner. Certification exams are administered independently by Microsoft and Anthropic and are not included in programme tuition. Exam details are accurate as of 21 August 2026; always confirm current format, pricing, validity and blueprint version with the issuer before booking.
Microsoft Foundry
GPT and Foundry Models
Python
