AI-103 Question Bank 2026: 25 Practice Questions for the Azure AI Apps & Agents Developer Exam

AI-103 (Azure AI Apps and Agents Developer Associate) is a developer exam and it expects you to build generative AI apps and agents on Azure. That makes practice questions more valuable here than at fundamentals level, because the exam tests decisions: which service, which parameter, which grounding strategy. This AI-103 question bank gives you 25 scenario questions with answers and explanations, mapped to the five skills-measured domains and weighted the way the real exam is. 

Attempt all 25 questions and review every explanation. Target 80%+ before booking the Azure AI Apps and Agents Developer exam.

 

Q1. Your app must block sexual, violent and self-harm content in both prompts and model output. Which service do you add?

A) Azure AI Search  

B) Azure AI Content Safety  

C) Azure Monitor  

D) Azure Key Vault

Answer: B — Azure AI Content Safety

Content Safety classifies and filters harmful content on input and output. Azure OpenAI also applies default content filters you can configure.

 

Q2. You want your deployed app to call Azure AI services without storing API keys. The recommended approach is:

A) Store the key in an environment variable on the client application

B) Microsoft Entra ID with a managed identity

C) Implement a custom JSON Web Token (JWT) generator

D) Authenticate using Azure AI Search vector embeddings

Answer: B — Managed identity with Entra ID

Managed identities remove secrets from code entirely; keys, when used, belong in Key Vault — never in source or URLs.

 

Q3. Which is the correct top-level resource for building, deploying, and governing generative AI apps and agents on Azure?

A) Azure AI Foundry project

B) Azure Machine Learning Workspace

C) Azure Cognitive Services Multi-Service Account

D) Azure Application Insights

Answer: A — Azure AI Foundry project

A Foundry project is where you deploy models, build agents, attach tools, evaluate and monitor.

 

Q4. You must track token usage and latency of your model deployment over time. You configure:

A) Diagnostic settings sending metrics/logs to Azure Monitor

B) Azure AI Model Catalog telemetry

C) Azure Cost Management budget alerts

D) Azure AI Content Safety usage logs

Answer: A — Diagnostic settings to Azure Monitor

Diagnostic settings route metrics and logs to Azure Monitor / Log Analytics for cost, usage and performance tracking.

 

Q5. Which responsible-AI step belongs in the plan phase of an AI solution?

A) Define an impact assessment and mitigations before building

B) Configure content filter severity levels in Azure OpenAI

C) Implement model reflection and self-critique loops

D) Deploy a shadow-model for live A/B testing

Answer: A — Impact assessment up front

Responsible AI is designed in, not bolted on. Identify harms and mitigations during planning, with human oversight for high-impact actions.

 

Q6. Before your code can call a GPT model in Azure OpenAI, you must first:

A) Create a model deployment

B) Provision a dedicated GPU cluster

C) Register a custom webhook in Azure AI Foundry

D) Upload training data to fine-tune the model

Answer: A — Create a model deployment

You deploy a base model under a deployment name, then reference that name in API calls. No training is required to consume it.

 

Q7. The most reliable place to set an agent's persistent behaviour and rules is the:

A) System (developer) message / instructions

B) Few-shot examples in the user prompt

C) The deployment configuration template

D) Azure AI Foundry orchestration settings

Answer: A — System message / instructions

The system message defines role, tone, and constraints for the whole conversation; per-turn user content should not carry standing rules.

 

Q8. A user reports the model is too random and inconsistent for a structured data-extraction task. You should:

A) Lower the temperature (e.g. to 0–0.2)

B) Increase the top_p value to 0.9

C) Increase the frequency_penalty parameter

D) Switch to an embedding model

Answer: A — Lower the temperature

Lower temperature makes output more deterministic — right for extraction and classification. High temperature suits creative generation.

 

Q9. Your long prompt plus document context is being truncated. The limiting factor is the model's:

A) Context window (token limit)

B) Max completions limit

C) Provisioned Throughput Unit (PTU) allocation

D) Azure Monitor log retention limit

Answer: A — Context window

Input + output tokens must fit the context window. Trim, summarise, or retrieve only relevant chunks (RAG) to stay within it.

 

Q10. You want the model to call your own getOrderStatus(orderId) API when a user asks about an order. You use:

A) Function/tool calling

B) Retrieval-Augmented Generation (RAG)

C) Fine-tuning with few-shot examples

D) Semantic Kernel memory connectors

Answer: A — Function / tool calling

Tool (function) calling lets the model request a structured call to your code; your app runs it and returns the result for grounding.

 

Q11. In the Azure AI Foundry Agent Service, a single execution of an agent over a conversation session is a:

A) Run

B) Thread

C) Flow

D) Trace

Answer: A — Run

The thread holds the conversation; a run is one execution of the agent over that thread; tools are the capabilities it can invoke.

 

Q12. A complex request spans billing, shipping and returns, each with its own specialised agent. The right pattern is:

A) Multi-agent orchestration with a coordinator and sub-agents

B) Chained sequence flows in Azure Logic Apps

C) A single unified fine-tuned model for all domains

D) Round-robin load balancing across model deployments

Answer: A — Multi-agent orchestration

An orchestrator delegates to domain sub-agents (increasingly via agent-to-agent protocols) and composes the result.

 

Q13. Users complain the assistant invents policy details. The most effective fix is to:

A) Ground responses on your policy documents via RAG

B) Adjust the presence_penalty to penalize new tokens

C) Fine-tune the model solely on the policy documents

D) Increase the model's temperature parameter

Answer: A — Grounding via RAG

Retrieval-augmented generation supplies trusted source text at query time, cutting hallucination far more reliably than prompt tweaks.

 

Q14. You need to adjust how strictly the Azure OpenAI deployment filters borderline content for an approved internal use case. You:

A) Configure the content filter / request modified filtering as allowed

B) Disable the system message metaprompt

C) Bypass Azure AI Content Safety using a private endpoint

D) Use a smaller base model with fewer safety guardrails

Answer: A — Configure content filtering

Content filter levels are configurable (with approval for relaxed settings); governance and logging stay on.

 

Q15. You have 10,000 labelled examples and need the model to adopt a very specific output style no prompt achieves. The right tool is:

A) Fine-tuning

B) Retrieval-Augmented Generation (RAG)

C) Multi-agent orchestration

D) Semantic ranking in Azure AI Search

Answer: A — Fine-tuning

Fine-tuning teaches a durable style/behaviour from many examples. Use RAG for facts/freshness; use fine-tuning for form — and prefer RAG first when unsure.

 

Q16. You need high-accuracy text extraction from scanned PDFs and photos. The right API is:

A) Azure AI Vision Read (OCR)

B) Azure AI Language Key Phrase Extraction

C) Azure AI Search vector search

D) Azure AI Content Safety image analysis

Answer: A — Azure AI Vision Read (OCR)

The Read API extracts printed and handwritten text; Document Intelligence adds structured field extraction for forms.

 

Q17. Prebuilt image tagging misses your factory's custom parts. You should:

A) Train a custom model in Azure AI Vision

B) Use the prebuilt Spatial Analysis container

C) Extract the metadata using Azure AI Language

D) Apply a generic bounding box filter

Answer: A — Train a custom model

Custom image classification/object detection is trained on your labelled images when prebuilt categories do not fit the domain.

 

Q18. You want one model to answer questions about the contents of an uploaded image in natural language. The best fit is:

A) A multimodal model (e.g. GPT-4o)

B) Azure AI Vision Read API combined with Translator

C) An embedding model (e.g. text-embedding-3)

D) A traditional Object Detection bounding-box model

Answer: A — Multimodal model

Multimodal models reason over image + text together, enabling visual question answering beyond plain OCR.

 

Q19. You must extract organisations and locations from support tickets. You use Azure AI Language's:

A) Named entity recognition

B) Key phrase extraction

C) Text summarization

D) Intent recognition

Answer: A — Named entity recognition

NER categorises entities such as organisations, locations and people from unstructured text.

 

Q20. Before sending customer text to a model, you must remove emails and phone numbers. You use:

A) PII detection and redaction in Azure AI Language

B) Azure OpenAI Content Filtering blocklists

C) Data Loss Prevention (DLP) policies in Azure Storage

D) Prompt engineering with strict formatting constraints

Answer: A — PII detection and redaction

Azure AI Language detects and redacts personally identifiable information — a common responsible-AI preprocessing step.

 

Q21. Your app must transcribe a live call and then translate it. The right pair of services is:

A) Azure AI Speech (speech-to-text) + Azure AI Translator

B) Azure AI Language Conversation Understanding + Azure AI Services

C) Azure OpenAI whisper model + Azure AI Content Safety

D) Azure AI Document Intelligence + Azure AI Speech

Answer: A — Speech-to-text + Translator

Speech transcribes audio; Translator converts the text across languages. Speech translation can also combine both.

 

Q22. To let the model retrieve semantically relevant passages from 50,000 documents, you build:

A) A vector index in Azure AI Search

B) A relational schema in Azure SQL Database

C) A caching layer in Azure Cache for Redis

D) A document classification model in Azure AI Language

Answer: A — Vector index in Azure AI Search

Azure AI Search stores embeddings and supports vector and hybrid search — the retrieval backbone of a RAG pipeline.

 

Q23. To create the vectors for that index, you call a(n):

A) Embedding model (e.g. text-embedding-3)

B) Autoregressive text generation model (e.g. GPT-4)

C) Named Entity Recognition (NER) pipeline

D) Semantic ranker endpoint

Answer: A — Embedding model

Embedding models turn text into numeric vectors so semantically similar content sits close together in vector space.

 

Q24. Retrieval quality is poor because whole 40-page documents are indexed as single records. The fix is:

A) Chunk documents into smaller passages before embedding

B) Increase the top_k parameter during vector search

C) Switch from vector search to keyword (BM25) search

D) Fine-tune the embedding model on the 40-page documents

Answer: A — Chunk the documents

Chunking into coherent passages makes retrieval precise and keeps retrieved context within the token budget.

 

Q25. The fastest way to ground an Azure OpenAI chat app on your own data without building a full pipeline is:

A) The "Add your data" (on-your-data) feature with Azure AI Search

B) Training a custom foundational LLM from scratch

C) Implementing a LangChain routing agent in Azure Functions

D) Storing all documents as system messages in a prompt template

Answer: A — The "Add your data" (on-your-data) feature with Azure AI Search

Azure OpenAI's on-your-data feature wires a search index into the chat completion so answers cite your content with minimal code.

Where to go from here

Our AI-103 study guide gives a week-by-week plan, the AI-103 certification guide covers scope and format and the AI-103 course on SkillTech maps each domain to hands-on labs. Build the agent patterns for Domain 2 with our step-by-step Foundry agent tutorial and see where AI-103 sits in the Azure AI Certification Roadmap.

Cross-check scope against the Azure AI Foundry documentation and Azure OpenAI documentation.

Frequently asked questions

Are these the real AI-103 exam questions?

No. These are AI-103 practice questions with explanations written to match the exam's style and scope. Reusing live items violates Microsoft's policy — practising the decisions is what raises your score.

How many questions are on the real AI-103 exam?

Typically 50–60 questions in about 100 minutes, with a passing score of 700 out of 1000.

Do I need to code for AI-103?

Yes — it is a developer associate exam. Python or C# experience is recommended and several questions assume you can read SDK-level logic.

Is AI-103 the replacement for AI-102?

AI-103 (Azure AI Apps and Agents Developer) is the current developer path, with far heavier emphasis on generative AI and agents than the retired AI-102.

Turn this AI-103 question bank into a verified credential with  SkillTech Club.