Exam Guide · Explained simply
Azure AI-103, explained like you're 12
Exam: AI-103 — Developing AI Apps and Agents on Azure
Badge you earn: Microsoft Certified: Azure AI Apps and Agents Developer Associate
Heads up: This exam replaced the old AI-102 exam, which retired on June 30, 2026. If you find old AI-102 material online, don't study from it — the new exam is about agents, generative AI, and Microsoft Foundry.
First, the big picture
Imagine you're building a super-smart robot helper. It can read, write, look at pictures, listen to people talk, and even use tools like a search engine or a calculator to get its job done.
Microsoft gives you a giant workshop full of parts and machines to build that robot. That workshop is called Microsoft Foundry (older name: "Azure AI Foundry" — same idea). The AI-103 exam checks whether you know how to use the workshop to build, launch, and take care of smart helpers.
The exam = 5 rooms in the workshop
1
Plan & manage an Azure AI solution25–30%
Pick the right tools, set them up safely, watch over them, keep them fair
2
Generative AI & agentic solutions30–35%
Build robots that write, chat, and use tools
3
Computer vision solutions10–15%
Make and understand pictures and videos
4
Text analysis solutions10–15%
Understand words, feelings in text, and speech
5
Information extraction solutions10–15%
Pull facts out of documents; give the robot a searchable memory
700 / 1000
to pass (that's 70%)
40–60
questions · 120 minutes
Words you'll hear a LOT
Learn these first — everything else builds on them.
Model
the "brain" that does the thinking.
LLM
a big brain that's great with words. Smart but pricier.
SLM
a smaller brain. Cheaper and faster, good for simple jobs.
Multimodal
a brain that understands more than one thing at once — words AND pictures AND sound.
Prompt
the instructions you type to the brain.
Token
a little chunk of text. You pay per token, like paying per minute on a phone plan.
Agent
a helper that doesn't just answer — it takes actions and uses tools.
Grounding
making the robot answer from real, trusted info you gave it, instead of guessing.
RAG
an open-book test: look up the right pages first, then answer using them.
Responsible AI
the safety rules that keep the robot kind, fair, and safe.
Room 1 · Core domain
25–30%
Plan & Manage an Azure AI Solution
Think of this room as being the boss of the project. Before you build, you plan. After you build, you take care of it.
1a. Picking the right tool for the job
Microsoft Foundry has many tools. Your job is to pick the right one, the same way you'd pick a hammer for a nail and a screwdriver for a screw.
Need to write or chat? → a language model (LLM or SLM).
Need to look at pictures? → a vision tool.
Need to understand speech? → a speech tool.
Need the robot to look things up? → a search / retrieval tool (vector search).
Building a robot helper that uses tools? → the Foundry Agent Service.
Exam tip: Pick the cheapest tool that still does the job. Don't use a giant LLM when a small model would work.
1b. Setting it up
Azure AI resource = renting your spot in the workshop.
Deploy a model = turn the brain on so your app can talk to it.
CI/CD pipeline = a conveyor belt that automatically tests and ships your app when you make changes.
1c. Watching over it (monitoring)
Performance — is it fast or slow?
Cost — are we spending too much (too many tokens)?
Drift — is the robot slowly getting worse or weirder over time?
Grounding quality — is it still using real facts, or starting to make things up?
1d. Keeping it safe (security)
Managed identity / keyless = the app proves who it is without carrying a password around. A special badge instead of a password on a sticky note.
Private networking = keep data on a private road instead of the public highway.
Role policies (RBAC) = give people the keys only to the rooms they actually need.
1e. Responsible AI (the safety rules)
Content filters / guardrails = block rude, harmful, or dangerous stuff.
Blocklists = a list of words or topics that are never allowed.
Prompt shields = protect the robot from trick questions that try to make it break its rules (a "prompt injection attack").
Content Safety = the Azure tool that checks text and images for unsafe content.
Auditing = a logbook of what the robot did, so you can check later.
Room 2 · Core domain · biggest
30–35%
Generative AI & Agentic Solutions
This is the heart of the exam. It's about building robots that create things and do tasks.
2a. Building generative apps
Deploy a model and send it prompts to get back writing, code, or answers.
RAG (open-book test) — the most important idea here:
- Store your documents in a searchable place.
- When a user asks something, the app searches for the most relevant pieces.
- It hands those pieces to the model along with the question.
- The model answers using those real pieces — fewer made-up answers.
Evaluate the app — test whether answers are good, relevant, and safe, and catch "fabrications" (also called hallucinations).
2b. Building agents (helpers that use tools)
An agent is more than a chatbot. It has:
Role & goal — what is its job?
Tools — search, APIs, your documents, custom functions.
Memory — so it remembers the conversation.
Function calling = the robot can press buttons / call functions to actually do things ("check the weather", "book a room").
Multi-agent solution = several robots working as a team, each with a job, with one "manager" robot organising them. Autonomous vs. semi-autonomous = does it act all by itself, or ask a human to approve first? For risky actions, add an approval step.
2c. Making it better and keeping it running
Prompt engineering = writing better instructions so you get better answers.
Temperature = a "creativity dial." Low = safe and predictable. High = creative but more likely to be wrong.
Reflection / self-critique = the robot checks its own work and fixes mistakes, like re-reading homework before turning it in.
Observability = tools to see inside what the robot is doing: tracing, token counts, safety signals, and latency.
Room 3 · Supporting domain
10–15%
Computer Vision Solutions
This room is about pictures and videos — both making them and understanding them.
3a. Making images and videos
Text-to-image = type "a red dragon on a mountain" and the model draws it.
Text-to-video = same idea, but a short video.
Inpainting / mask edits = erase part of a picture and let the AI fill it in. ("Mask" = the shape you selected to change.)
3b. Understanding images and videos
Captioning = the AI writes a sentence describing a picture.
Visual question-answering = show a picture, ask "how many apples?", it answers from the picture.
Alt-text = a description of an image for people who can't see it (accessibility).
Content Understanding = a Foundry tool that pulls useful details out of images and video.
Object detection = finding where things are (drawing a box around each cat).
3c. Safety for pictures
Filters block unsafe or banned images.
Indirect prompt injection in images = someone hides sneaky text inside a picture to trick the robot. Detect and block it.
Watermarks & brand rules = mark AI images and keep logos used correctly.
Room 4 · Supporting domain
10–15%
Text Analysis Solutions
This room is about understanding words and speech.
4a. Understanding text
Entity extraction = pulling out the important nouns — names, places, dates, companies.
Summarization = making a long thing short.
Structured JSON output = a neat, organised box a computer can use, instead of messy free text.
Sentiment / tone = is the text happy, sad, or angry? Great for reviews.
Translation = one language to another, using Azure Translator or an LLM.
4b. Speech
Speech-to-text = the robot listens and types out what it heard.
Text-to-speech = the robot reads out loud.
Custom speech models = teaching it special words (your product names) so it hears them right.
Speech translation = listen in English, speak back in Spanish.
Room 5 · Supporting domain
10–15%
Information Extraction Solutions
This room gives the robot a great memory and reading skills — so it can pull facts out of documents and look them up later.
5a. Retrieval and grounding (the robot's library)
Ingest and index = load your documents in and organise them so they can be searched fast.
Semantic search = search by meaning, not exact words. "car" also finds "automobile."
Vector search = turns words into number-fingerprints (vectors) and finds things that mean the same. This powers good RAG.
Hybrid search = uses both keyword and vector search together for the best results.
OCR = the robot reads text out of a picture or scanned page.
5b. Reading documents
Document Intelligence = pulls info out of forms, receipts, invoices — knowing "this is the total, this is the date."
Layout analysis = understands where things sit on the page (tables, columns, headings).
Field extraction = grabs specific values ("Invoice Number: 4471").
Clean, grounded output = turns messy documents into neat text an agent or RAG app can use.
How it all connects
You use Microsoft Foundry (Room 1) to pick and set up the right brains and tools safely. You build generative apps and agents (Room 2) that create things and do tasks, and use RAG so they stay truthful. You add vision (Room 3), text and speech (Room 4), and information extraction (Room 5) to give the robot a searchable memory. The whole time you follow Responsible AI rules and watch cost, speed, and safety.
Quick check before the exam
I can explain RAG and why it stops hallucinations.
I know the difference between LLM, SLM, and multimodal — and when to pick each.
I understand what an agent is and how function calling, tools, and memory work.
I know managed identity / keyless auth is safer than passing keys around.
I can name Responsible AI tools: content filters, blocklists, prompt shields, Content Safety.
I understand vector search, semantic search, and hybrid search.
I know OCR, Document Intelligence, and Content Understanding are for reading documents.
I know the temperature dial controls creativity vs. predictability.
I know the pass mark is 700/1000 (70%).
Source of exam structure: Microsoft Learn — "Study guide for Exam AI-103: Developing AI Apps and Agents on Azure" (skills measured as of April 16, 2026). Always double-check the live study guide before your exam date, because Microsoft updates objectives from time to time.
