How many free AI-103 practice questions can I try?

You can answer 10 AI-103 practice questions free without creating an account. Each one is graded instantly with a written explanation. A free account keeps your progress, and the full bank (300+ questions) unlocks with a subscription.

What topics does the AI-103 exam cover?

The AI-103 exam is organised into 5 domains: Implement generative AI and agentic solutions (33%), Plan and manage an Azure AI solution (27%), Implement information extraction solutions (14%), Implement computer vision solutions (13%), Implement text analysis solutions (13%). The weighting above shows how much of the exam each domain represents, so budget study time in the same proportion.

How long is the AI-103 exam and how is it scored?

The AI-103 exam has approximately 50 questions in 120 minutes, with a passing score of 700. Practice sets here are scored the same way as the real exam so you can see whether you are above or below the line before test day.

How should I start studying for Azure AI Apps and Agents Developer (AI-103)?

Start with Implement generative AI and agentic solutions, the highest-weighted domain, then work down the list. Answer a mixed set after each domain pass — mixing domains forces you to identify what a question is asking before recalling the content, which is what the real exam does.

Are these the same questions that appear on the real AI-103 exam?

No. These are original practice questions written against the published AI-103 exam objectives. They mirror the style, difficulty, and domain weighting of the real exam, but the actual exam content is confidential and is not reproduced here.

AI-103 Practice Questions — Azure AI Apps and Agents

Domain	Exam weight	What it covers
Implement generative AI and agentic solutions	33%	Build generative apps, RAG, Foundry agents, tools, memory, multi-agent orchestration, and gen-AI ops
Plan and manage an Azure AI solution	27%	Choose Foundry services and models, set up deployments, secure and monitor solutions, apply responsible AI
Implement information extraction solutions	14%	Retrieval pipelines, vector/hybrid search, enrichment, and document extraction
Implement computer vision solutions	13%	Image and video generation, multimodal understanding, provenance and safety
Implement text analysis solutions	13%	Entities, sentiment, translation, and speech as an agent modality

Sample question 1 — Implement text analysis solutions

A biotech team must pull drug-trial identifiers and internal compound codes out of research abstracts. These terms are not recognized by the prebuilt entity types, and the team has a few hundred labeled examples. Which Language capability should the team use?

A. Train a custom named entity recognition model on the labeled examples to extract the domain terms.
B. Use the prebuilt entity recognition model and keep only the entities that look like trial identifiers.
C. Train a custom text classification model on the abstracts to label each abstract by trial phase.
D. Fine-tune a summarization model on the abstracts and read the identifiers out of each summary.

Answer: A. Option A is correct because a custom named entity recognition model learns the team's own entity types from labeled examples and extracts drug-trial identifiers and compound codes. Option B is incorrect because the prebuilt model does not know the team's private entity types, so filtering cannot recover them. Option C is incorrect because text classification labels the whole abstract and does not extract the identifiers. Option D is incorrect because a summary describes the abstract and does not reliably preserve exact codes.

Sample question 2 — Implement generative AI and agentic solutions

A team ships a RAG assistant and needs, for every production conversation, which passages were retrieved and whether the answer was grounded in them, and it must compare two model versions. The check must not slow the live response. What should the team do?

A. Add a synchronous groundedness check inside the live request path so that every answer is scored before it returns to the user.
B. Capture retrieved passages and answers in traces, then score the two model versions offline.
C. Review a random sample of answers by hand every week and record each one that appears to be grounded incorrectly.
D. Retrieve more passages for every production conversation so the model can answer with more context each time.

Answer: B. Option B is correct because tracing stores the retrieved passages and the answer for each conversation, and offline evaluators score groundedness and compare the two model versions without touching the live path. Option A is incorrect because a synchronous check adds latency to every request. Option C is incorrect because manual sampling cannot attribute grounding per conversation at scale. Option D is incorrect because more passages can reduce groundedness and does not produce a comparable metric.

Sample question 3 — Plan and manage an Azure AI solution

A public health chatbot must keep self-harm content blocked at the strict level, but benign questions about medication dosage are rejected because they trip the violence category of the default content filter. The team wants the medical questions to pass without weakening self-harm blocking. What should the team configure?

A. A custom filter policy that sets a per-category severity threshold so the over-blocking category is tuned.
B. A prompt-filter setting that turns harm screening off for prompts and relies on the completion filter.
C. A custom blocklist that adds the medical terms so the filter stops flagging the dosage questions.
D. A second model that labels each message as medical before the content filter runs on it.

Answer: A. Option A is correct because a custom content filter sets a severity threshold per category, so self-harm stays at its configured strict level while the violence threshold that rejects medical wording is tuned. Option B is incorrect because turning off the prompt filter removes harm screening for every category. Option C is incorrect because a blocklist adds terms to block and does not exempt medical wording from a category. Option D is incorrect because pre-classifying every message adds latency without setting per-category severity levels.

Azure AI Apps and Agents Developer (AI-103) Practice Questions

Azure AI Apps and Agents Developer (AI-103) Practice Questions

How the AI-103 exam is structured

Worked sample AI-103 questions