Free
- Untimed practice across all five domains
- Every question explained, with its domain and subskill named
Microsoft · AI-103 · Azure AI Apps and Agents Developer Associate
AI-102 is retired. AI-103 isn't a rename.
Microsoft retired AI-102 on 30 June 2026 and rebuilt the exam around Microsoft Foundry, the agent runtime, and shipping generative AI rather than wiring up pre-built services. Generative AI and agentic solutions is now the largest domain. Our practice is written to the current skills outline, not to AI-102 material with the names changed.
- Questions weighted to the official domain split, so a mock matches the real mix
- Every item tagged to the domain and subskill it tests
- Readiness shown against 700, not as an invented percentage
Free practice across all five domains. No card required.
AI-103Implement generative AI and agentic solutionsHard
A retailer is shipping a customer-facing agent from a Microsoft Foundry project. The agent must answer from a product catalogue already indexed for retrieval, call an internal order-status API, and stop for a human before it issues a refund.
Which two actions should you take? Select two.
Two choices only. Your earliest pick was dropped.
0 of 2 selected
An original item written against the skills outline dated 16 April 2026. Not a real exam question.
AI-102 to AI-103
The centre of gravity moved to agents
AI-102 tested whether you could deploy and integrate Azure's pre-built AI services. AI-103 tests whether you can build and ship a generative application or an agent on Microsoft Foundry. Planning, vision, language and knowledge mining carry over, but the weighting changed and the largest domain is now one that barely existed on the old exam.
If you prepared for AI-102, most of your platform knowledge still counts. What doesn't carry is the assumption that the exam is service-by-service.
| AI-102 | AI-103 | |
|---|---|---|
| Status | Retired 30 Jun 2026 | Current |
| Credential | Azure AI Engineer Associate | Azure AI Apps and Agents Developer Associate |
| Centre of gravity | Deploying pre-built AI services | Building apps and agents on Foundry |
| Largest domain | Service implementation | Generative AI and agentic solutions |
AI-102 is retired. The current associate credential for this work is Azure AI Apps and Agents Developer Associate, earned by passing AI-103.
Scoring
700 is a pass. It isn't 70%.
Microsoft reports technical exams on a scale of 1 to 1,000, and 700 or above passes. Because it's a scaled score, it doesn't correspond to answering 70% of questions correctly — how many you need depends on the difficulty of the questions you're given.
There's no penalty for a wrong answer, so never leave one blank. Some questions are worth more than one point, and on some exams you can't return to a question once you've moved past it.
How Microsoft scales a score1–1,000 · 700 to pass
You answered
72%of the questions correctly
700
Your scaled score lands somewhere around670–730Could go either way
Now change nothing except which form you were given
Same answers. Harder form, higher scaled score — that's what scaling is for. It's also why nobody can tell you the percentage you need.
Positions here are illustrative. Microsoft publishes the 1–1,000 scale and the 700 pass mark, but not the conversion from raw answers — which is the point. When a number can't be known exactly, a range is the honest way to report it.
The loop
Practice, sit, review, repeat
Step 1 is free. Steps 2 to 4 need a subscription.
certs.acedify.com/ai-103
Practice · untimed
HardA retailer is shipping a customer-facing agent from a Microsoft Foundry project. The agent must answer from a product catalogue already indexed for retrieval, call an internal order-status API, and stop for a human before it issues a refund.
Which two actions should you take? Select two.
Define tool schemas for catalogue retrieval and the order-status API, and enable function calling on the agent.
Paste the full catalogue into the system prompt and raise max_tokens so every SKU fits.
Fine-tune a vision model on product photos so the agent can skip retrieval altogether.
Add an approval-flow control on the refund action so the agent cannot execute it autonomously.
Tool schemas and function calling connect the agent to retrieval and the order-status API. An approval-flow control is how you stop a refund from running autonomously. Prompt stuffing and vision fine-tuning do not. Domain: Implement generative AI and agentic solutions.
Click a step to see it.
What you get back
Not a prediction. A distance from the line.
Your readiness is a range, because it's an inference from the questions you've answered rather than a result you've earned. For a certification that matters more than it does for a scored test, because the outcome is binary: the only question is whether the whole range sits above 700.
While the range straddles the line, you're a coin flip, and we'll say so.
AI-103 · 4 mocks completed
Where you stand
Borderline
530–720
Pass mark 700
Your range crosses the pass mark. On today's evidence it's a coin flip.
Three states
Try the three states. We won't show you a percentage chance of passing — nobody can calculate one honestly without years of real outcomes to calibrate against.
We won't show you a percentage chance of passing. Nobody can calculate one honestly without years of real pass and fail outcomes behind it.
Domains
What AI-103 actually covers
Microsoft publishes a weighted skills outline. Our mocks follow those weights, so a domain worth a third of the exam gets a third of the questions.
| Domain | Weight |
|---|---|
| Implement generative AI and agentic solutions | 30–35% |
| Plan and manage an Azure AI solution | 25–30% |
| Implement computer vision solutions | 10–15% |
| Implement text analysis solutions | 10–15% |
| Implement information extraction solutions | 10–15% |
Weights are ranges because Microsoft publishes them as ranges. Skills measured as of 16 April 2026. Microsoft revises outlines, and we re-check ours against theirs on a schedule — the date we last looked is in the claims register.
The first two domains are more than half the exam. If you're budgeting study time, two thirds of it belongs in Foundry and agents.
Information extraction is one of the lighter domains on paper, but grounding and retrieval run through the agent questions too. It's tested in more places than its weight suggests.
What the heaviest domain contains
Connecting an application to a Foundry project, the agent runtime, agent tools and function calling, multi-agent orchestration, plus tuning, evaluation, tracing and observability. Planning also asks which Foundry service fits a constraint, not what each service does.
Exam facts
Prices vary by country. Check Microsoft Learn for yours.
- Exam
- AI-103
- Credential
- Azure AI Apps and Agents Developer Associate
- Length
- 120 minutes
- Pass mark
- 700 on a 1–1,000 scale
- Delivery
- Pearson VUE
- Skills measured as of
- 16 April 2026
Pricing
Practice free. Pay when you want to know where you stand.
Pro
- Timed mocks at exam length
- A scaled score against the 700 pass mark
- Domain-weighted readiness, and your range over time
- Everything on every exam, one subscription
One subscription covers the Microsoft certifications, AWS, the Claude certifications, TOEFL, IELTS and the SAT.
The exam fee is separate and goes to Microsoft. We're prep, not registration.
Start practicing for freeFAQ
AI-103 questions
Find out which domain is keeping you under 700.
Free practice across all five domains, no card. The timed mocks and the readiness range are there when you want them.
Start practicing for freeAlso available for AI-901, AI-200, AI-300, AZ-104 and the Claude certifications on the same subscription.