AIAI Agent Dev Labaiagentdev-lab.com
Sign in to save progressSign in
Menu

STEP 1 · Study

AI-103 StudyAzure AI Apps and Agents Developer

Read the key points for each exam domain, then use “Practice this domain” to answer questions from that domain only. Domains with a larger weight appear more often on the exam.

01

Plan and manage an Azure AI solution

25–30%

Learning goals

  • Choose LLMs, small language models, multimodal models and Foundry Tools to fit the task
  • Configure security with managed identities, private networking, keyless authentication and RBAC
  • Manage quotas, rate limits and cost, and monitor quality and safety
  • Implement responsible AI with guardrails, evaluation, tracing and approval flows

Key points

  • Model choice balances quality, latency and cost: small language models for high-volume routine classification, large models for complex reasoning, multimodal when images are involved.
  • Keyless authentication is the rule: assign a managed identity to the app and grant only the minimum roles with RBAC. Never put API keys in code or repositories.
  • Isolate the network with private endpoints and similar, and narrow public exposure.
  • Handle 429 (rate limit) with quota checks, retries with exponential backoff and capacity planning. Cost is driven by token volume and deployment type.
  • Monitor not only model performance but also grounding quality, safety events and the health of search indexes.
  • Responsible AI: design content filters and guardrails, evaluation, tracing and provenance metadata, approval workflows and tool access control as a set.

Workplace uses

  • SalesKeep costs down by splitting models by use: a large model for proposal drafts and a small model for inquiry classification.
  • General affairsBuild the internal policy Q&A app with a managed identity and a private network, removing API key management and leak risk.
  • FinancePayment tools are never run directly by the agent but go through an approval workflow. Tracing records who approved what, on what grounds.

Exam objectives (from the exam guide)

  • Choose the appropriate Foundry services for generative AI and agents
  • Set up AI solutions in Foundry
  • Manage, monitor, and secure AI systems
  • Implement responsible AI across generative AI and agentic systems

Learning resources

Learn in detail (4 topics with mini quizzes) →Practice this domain (85 questions) →Next domain ↓
02

Implement generative AI and agentic solutions

30–35%

Learning goals

  • Implement generative apps with RAG and tool integration using the Foundry SDK
  • Build agents with defined roles, goals, tool schemas and conversation management
  • Design multi-agent and semi-autonomous workflows with approvals
  • Optimize operations with generation parameter tuning, tracing and token analysis

Key points

  • RAG collects grounding with search and passes it to the model to answer, reflecting internal information without retraining the model.
  • Define agent tools clearly with names, descriptions and parameter types. Vague descriptions cause the wrong tool to be called.
  • Continuing a conversation requires managing conversation history (threads). Decide what to remember and what to discard.
  • Multi-agent systems split roles and an orchestrator coordinates them. Do not cram everything into one agent.
  • High-impact operations such as ordering or payments must not run autonomously; require an approval flow.
  • Evaluate both quality (relevance, groundedness) and safety. Build in evaluations that detect fabrication (ungrounded answers).
  • Lower temperature and similar settings for work that needs consistency. Ensure observability with tracing, token analysis and latency breakdowns.

Workplace uses

  • SalesReference past meeting notes per customer with RAG and draft talking points for the next visit; the account owner reviews the final proposal.
  • General affairsSupply ordering agent: separate the inventory check tool from the ordering tool, and run orders only after approval.
  • FinanceMake invoice matching multi-agent (extraction, matching and exception reporting), with people deciding on discrepancies.

Exam objectives (from the exam guide)

  • Build generative applications by using Foundry
  • Build agents by using Foundry
  • Optimize and operationalize generative AI systems

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (99 questions) →Next domain ↓
03

Implement computer vision solutions

10–15%

Learning goals

  • Implement image and video generation and editing (such as inpainting) from text and reference images
  • Analyze images and video with multimodal models and Content Understanding
  • Address multimodal-specific risks such as indirect prompt injection through text in images

Key points

  • Image generation can use reference images in addition to text prompts. To fix only part of an image, use inpainting with a mask.
  • Question answering, captions and alt-text generation based on image content are strengths of multimodal models. Align alt text with accessibility requirements.
  • Use Content Understanding to extract structured features from images and video. Video is processed by segment.
  • Beware of indirect prompt injection, where text embedded in images is interpreted as instructions. Design so that image-derived text is never treated as instructions.
  • Apply policies to generated content such as watermarking, detection of prohibited content and brand rules.

Workplace uses

  • SalesGenerate catalog descriptions and alt text from product images; staff review the wording before publishing.
  • General affairsApply brand guidelines and watermarks to images generated for internal notices.
  • FinanceWhen reading receipt images, design so text in the image is never executed as instructions, preventing injection.

Exam objectives (from the exam guide)

  • Design and implement image- and video-generation solutions
  • Design and implement multimodal understanding workflows
  • Implement responsible AI for multimodal content

Learning resources

Learn in detail (3 topics with mini quizzes) →Practice this domain (38 questions) →Next domain ↓
04

Implement text analysis solutions

10–15%

Learning goals

  • Extract entities, summaries and structured JSON with generative prompts and Foundry Tools
  • Implement detection of sentiment, tone and sensitive information, and translation
  • Build speech recognition, speech synthesis and speech translation into agents

Key points

  • Extraction results feed later steps, so specify structured output such as JSON with a schema.
  • Detecting sentiment, tone, safety issues and sensitive information (such as personal data) directly supports inquiry routing and information management.
  • Choose between Azure Translator and LLM translation by use. When terminology must be consistent, consider dictionaries and glossaries.
  • When recognition of industry or internal terms is poor, improve it with a custom speech model.
  • Speech is one of an agent’s input and output modalities. Design the flow as speech → text → reasoning → speech.

Workplace uses

  • SalesTranscribe meeting recordings and extract key points, next actions and competitor names as JSON into CRM drafts.
  • General affairsDetect personal information in internal inquiries and mask it before recording.
  • FinanceTranslate English billing terms from overseas suppliers and unify terms such as account titles with an internal glossary.

Exam objectives (from the exam guide)

  • Apply language model text analysis
  • Implement speech solutions

Learning resources

Learn in detail (2 topics with mini quizzes) →Practice this domain (38 questions) →Next domain ↓
05

Implement information extraction solutions

10–15%

Learning goals

  • Ingest and index documents, images, audio and video
  • Configure semantic, hybrid and vector search
  • Extract structured data from documents with OCR, layout analysis and field extraction

Key points

  • Hybrid search combines keyword and vector search. It handles both model numbers and proper nouns and paraphrases.
  • Semantic ranking reorders search results by meaning and improves the relevance of the top results.
  • Scanned PDFs need an ingestion flow that includes OCR. Layout analysis preserves the structure of tables and headings.
  • Use enrichment (skills) to add text, image and layout information before indexing.
  • Producing structured output such as Markdown with Content Understanding makes it easier for later RAG and agent steps to handle.
  • Search pipelines can be connected directly as agent tools.

Workplace uses

  • SalesSearch across past proposals and quotes with hybrid search so both model-number queries and vague questions work.
  • General affairsIngest a printed policy book with OCR and layout analysis and use it as Q&A grounding while keeping its chapter structure.
  • FinanceExtract supplier, amount and payment due date fields from invoices and pass them to the matching agent.

Exam objectives (from the exam guide)

  • Build retrieval and grounding pipelines
  • Extract content from documents

Learning resources

Learn in detail (2 topics with mini quizzes) →Practice this domain (38 questions) →Try the mock exam →

When you finish reading

Practice all domains (298 questions)Mock exam (50 questions · 75 min)AI-103 home
Notes (unofficial, original questions, no pass guarantee)