AI-103 Azure Azure AI Apps and Agents Developer – Associate
- Vendor
- Azure
- Level
- Associate
- Questions
- 50
- Time
- 120 min
- Domains
- 5
- In the bank
- 303
Practice 303 questionsTake a timed mock examDay-by-day study plan
Guides for this exam
- Prompting, RAG, Fine-Tuning or Continued Pre-Training: One Rule That Works on AWS, Azure, Google Cloud and Databricks Exams · 한국어
- Seven Mistakes That Repeat Across the Azure Exams, from AZ-900 to AI-103 · 한국어
- AI-103 Study Guide: Follow One Request Through a Foundry App, Hop by Hop · 한국어
The exam at a glance
| Level | Associate |
|---|---|
| Questions | 50 |
| Time | 120 minutes (about 2.4 minutes per question) |
| Passing mark | 700 on a scaled score of 100–1,000 |
Domains and weights
| Domain | Weight | Practice questions here |
|---|---|---|
| D1 · Plan and manage an Azure AI solution | 27.5% | 87 |
| D2 · Implement generative AI and agentic solutions | 32.5% | 100 |
| D3 · Implement computer vision solutions | 12.5% | 38 |
| D4 · Implement text analysis solutions | 12.5% | 38 |
| D5 · Implement information extraction solutions | 12.5% | 40 |
54 of the 303 practice questions (18%) ask for more than one answer.
D1 · Plan and manage an Azure AI solution 27.5%
Planning questions start from an organizational constraint: two teams sharing one Foundry resource, a compliance rule about where agent data lives, or a ban on API keys. You decide how resources, projects, identities and guardrails should be arranged. Wrong options often try to fix the problem with a setting that sits at the wrong scope or cannot be changed after creation.
What to watch for
- Before choosing a topology, check whether the API in question works per project or only at the Foundry resource level. Isolation needs differ between the two.
- Some choices, such as basic versus standard agent setup, are made when a project is created. If the scenario needs to change one later, expect a rebuild rather than an update.
- For keyless access, two things must line up: the target service has to accept role-based access, and the caller's identity needs the narrowest data role that works.
Tasks in the official exam guide (4)
D2 · Implement generative AI and agentic solutions 32.5%
Most items here describe code that already exists and has a problem: a schema that is rejected, an endpoint read from a hand-edited file, a feature that stopped working, or an evaluator that scores fair answers too low. The task is to pick the supported fix in the current Foundry toolset. Distractors are usually older APIs, retired features or tools that measure the wrong thing.
What to watch for
- Strict structured outputs are picky about the schema itself. Check how required fields, optional fields and extra keys are declared on every object.
- Notice the status words in a scenario (retired, successor, GA only). They often decide between a familiar older approach and the one the current platform supports.
- Match the evaluator to what counts as correct: word-overlap metrics and LLM-judge metrics such as Similarity or Groundedness reward very different answers.
Tasks in the official exam guide (3)
D3 · Implement computer vision solutions 12.5%
Vision scenarios describe an image request that fails or gives weak results: a size the model rejects, small text misread, a private image the model cannot open, or alt text that ignores the page. You choose the parameter, input format or service that fixes it. The right answer usually respects a documented input rule; wrong answers change something unrelated or weaken privacy.
What to watch for
- When a model cannot open an image, think about where the image lives. Private storage calls for the app to fetch the file and send it inline as base64.
- Separate built-in harm categories from business-specific visual rules. Content Safety covers the first; a Content Understanding analyzer handles custom checks such as a banned logo.
- Image detail and context both change output quality: high detail helps with fine print, and page context helps the model describe why an image is there.
Tasks in the official exam guide (3)
D4 · Implement text analysis solutions 12.5%
Speech and language items describe a product complaint or a compliance request: a voice agent that cuts users off, PII that must be masked selectively, long narration that times out, or speech costs that look too high. You choose the mode, parameter or Azure Language feature that fits. The wrong options are usually real features built for a different volume, latency or level of control.
What to watch for
- Sort audio jobs by timing first: live conversation, short files and large stored batches each point to a different transcription or synthesis mode.
- For PII, read exactly which entity types must be hidden and which must stay visible. Azure Language lets the request limit detection to chosen categories.
- Prefer a core, non-generative Azure Language feature when the task has a fixed output, such as health entities, and no training data exists.
Tasks in the official exam guide (2)
D5 · Implement information extraction solutions 12.5%
Extraction questions follow the path from raw files to grounded answers: which Content Understanding analyzer suits a document set, how fields report confidence and source, and how Azure AI Search indexes and queries vectors. You decide on analyzers, settings and index parts. Many wrong answers confuse indexing-time steps with query-time steps, or rely on preview features when only GA is allowed.
What to watch for
- Pick the lightest analyzer that returns what the scenario needs. OCR and layout alone need no language model, while field extraction and RAG-ready output do.
- Keep the two halves of integrated vectorization apart: skills embed content while indexing, and a vectorizer on the vector profile embeds text queries.
- When review routing or source highlighting is required, look for the setting that makes the analyzer return per-field confidence and location.
Tasks in the official exam guide (2)
A study order that works
The 20-day study plan for AI-103 takes the domains in this order. Each day has a short reading, key terms and a quiz from the bank.
- Days 1–5: D1 · Plan and manage an Azure AI solution
- Days 6–11: D2 · Implement generative AI and agentic solutions
- Days 12–14: D3 · Implement computer vision solutions
- Days 15–16: D4 · Implement text analysis solutions
- Days 17–19: D5 · Implement information extraction solutions
- Day 20: Final Review: Exam Format, Commonly Confused Pairs, and Test-Day Plan
Recent changes to these questions
- : 3 answer choice sets changed, 3 answer keys corrected, 3 explanations rewritten
- : 60 questions added
- : 1 answer choice set changed, 1 answer key corrected, 1 explanation rewritten
About these questions
Every question is original, written from the public exam guide and checked against the vendor's own documentation. None comes from a real exam. Each one has an explanation of why the answer is right and why the other options are wrong. How the questions are written.