Welcome to AI-103: know the exam before you study it
When you join a new codebase, you read the README and the CI pipeline before you touch the code. They tell you what "done" means. The study guide does the same job for AI-103. It lists exactly what Microsoft will test, and this course follows it.
Ground rules first:
- The study guide is Microsoft's document. It is free on Microsoft Learn (the short link is
aka.ms/AI103-StudyGuide). This course teaches from it but does not reproduce it, so keep it open in a tab. - This course is independent of Microsoft and not endorsed by it.
- No pass is promised. Passing depends on your preparation and hands-on practice.
The exam is AI-103: Developing AI Apps and Agents on Azure. Passing it earns Microsoft Certified: Azure AI Apps and Agents Developer Associate. We study against the skills measured as of April 16, 2026. The guide describes the candidate as an Azure AI engineer who builds, manages and deploys agents and AI solutions using Microsoft Foundry, with experience developing apps in Python.
The five weighted skill areas
| Skill area | Weight |
|---|---|
| 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% |
The first two areas together make up 55–65% of the exam. That is where most of your study time should go. Here is the same idea as code, which is also how this course's 50 levels were planned:
# Share out the 48 content levels by the midpoint of each weight.
# Level 1 is orientation and level 50 is a final review, so they sit outside this.
weights = {
"Plan and manage": 27.5, # midpoint of 25-30%
"Generative AI & agents": 32.5, # midpoint of 30-35%
"Computer vision": 12.5,
"Text analysis": 12.5,
"Information extraction": 12.5,
}
total = sum(weights.values()) # 97.5, because midpoints don't add to exactly 100
for area, w in weights.items():
print(f"{area:24} ~{48 * w / total:4.1f} levels")
# Output:
# Plan and manage ~13.5 levels
# Generative AI & agents ~16.0 levels
# Computer vision ~ 6.2 levels
# Text analysis ~ 6.2 levels
# Information extraction ~ 6.2 levels
The course is laid out roughly like this: L2–L15 plan and manage, L16–L31 generative AI and agents, L32–L37 computer vision, L38–L43 text and speech, L44–L49 retrieval and document extraction, and L50 a final review. Use the same proportions for your own study hours.
Gotcha: the guide says its bullets illustrate how a skill is assessed, and that "related topics may be covered." Treat each bullet as the centre of a topic, not its edge.
Logistics: check the official pages, not forums
Exam details change. Before you book, go to the exam page on Microsoft Learn and check the current length, price and languages. At the time of writing, the page lists 120 minutes, a proctored exam, and ten languages, and it says price depends on the country where you take the exam. Things that are stable and worth knowing:
- Pass score: 700 or greater on a 1–1,000 scale. This is a scaled score, so it may not equal 70% of the points.
- No penalty for guessing. Never leave a question blank. Some questions are unscored, but you can't tell which ones, so answer every one as if it counts.
- Practice Assessment: it is on AI Skills Navigator, and you must be signed in there to launch it. There is also an exam sandbox where you can try the question interface.
- Language: if the exam isn't offered in your preferred language, you can request an extra 30 minutes. Localized versions are updated about eight weeks after the English one.
- Renewal: associate certifications expire annually. You renew with a free online assessment on Microsoft Learn.
Exam craft: three question shapes
- Scenario "best answer": a company, a goal and some constraints. Several options might work, but only one fits every constraint best.
- Product contrast: two or more similar Azure or Foundry services, and you pick the one that fits. Learn what makes each product different, not only what it does.
- Recognition: you identify a feature, setting or code snippet from its description.
The elimination routine
Run this on every scenario question:
1. Read the LAST sentence first: what exactly is being asked?
2. Mark the qualifier: "least effort", "minimize cost", "most secure", "fewest changes".
3. Cross out options that break a hard requirement (wrong region, wrong data type, no Python support...).
4. GA check: the guide says most questions cover GA features. Preview features appear
only if they are commonly used. If a GA option meets the need, prefer it.
5. Apply the qualifier to whatever options are left. That decides the answer.
Worked example. An app must call a model in a Foundry project. The security team requires the most secure approach, with no secrets to store or rotate.
- API key in an environment variable: it works, but it is a secret. Out.
- API key in Key Vault: better, but still a key you have to rotate. Out under "no secrets".
- Managed identity with Microsoft Entra ID (keyless): no secret exists at all. Best.
- Key hard-coded in source: it works, and it is the worst choice here.
If you swap the qualifier for "least development effort for a quick prototype", the key-based option may become the intended answer. The qualifier decides the answer. The guide lists "managed identity, private networking, keyless credentials, and role policies" under security, so expect this contrast to come up.
Trade-off reminder: in real projects, "least effort" and "most secure" often conflict. On the exam they are deliberate signals. Don't answer with what you'd do at work; answer the constraint as written.
What the exam asks about this
No exam item tests this orientation directly. What it gives you is: knowing the weights tells you where questions cluster, and the elimination routine plus the GA-first rule decide close calls in every section.
Key takeaways
- Five areas: plan and manage (25–30%), generative AI and agents (30–35%), and vision, text, and information extraction (10–15% each).
- Study against the skills measured as of April 16, 2026. Read the free guide on Microsoft Learn, and check length, price and languages on the exam page.
- 700 is a scaled pass mark, not 70%. Guessing costs nothing.
- The Practice Assessment is on AI Skills Navigator, and you must be signed in.
- Elimination routine: read the ask, mark the qualifier, cut options that break a requirement, prefer GA, then let the qualifier decide.
- This course is independent and not endorsed by Microsoft, and no pass is promised.