Prepare for Microsoft AI-103 with this 2026 Azure AI Apps and Agents Developer study guide. It maps the official Microsoft Learn exam skills to practical study notes, service-selection guidance, exam traps, and hands-on revision tasks for Microsoft Foundry, generative AI, agents, RAG, vision, speech, text analysis, and content understanding.
AI-103
Azure AI Apps and Agents Developer
Checked: 27 June 2026
Skills measured: 16 April 2026
Microsoft Learn mapped
AI-103 Study Guide 2026: Developing AI Apps and Agents on Azure
A practical Microsoft AI-103 exam guide for developers and Azure AI engineers. It maps the official skills measured to Microsoft Learn resources and explains what to focus on for Microsoft Foundry, model deployments, generative AI apps, agents, retrieval-augmented generation, multimodal AI, responsible AI, monitoring, and information extraction.
Use this as a companion to the official Microsoft Learn study guide: start with the domain overview, review the common exam traps, then use the objective tables to study each official bullet.
Official AI-103 exam scope
AI-103 is aimed at software engineers and Azure AI practitioners who build, manage, and deploy AI apps and agents using Microsoft Foundry. You should be comfortable with Python, APIs, SDKs, generative AI concepts, general AI services, Azure identity, and production application patterns.
Plan and manage an Azure AI solution25–30%
Implement generative AI and agentic solutions30–35%
Implement computer vision solutions10–15%
Implement text analysis solutions10–15%
Implement information extraction solutions10–15%
Exam reality: AI-103 is not only about naming services. Expect scenario questions where you choose the correct model, Foundry service, retrieval strategy, agent tool, deployment option, safety control, monitoring approach, or multimodal extraction pattern.
Primary source: Microsoft Learn AI-103 study guide · Azure AI Apps and Agents Developer Associate certification · Course AI-103T00
Jump to section
How to use this guide Common exam traps 1. Plan and manage 2. Generative AI and agents 3. Computer vision 4. Text analysis 5. Information extraction Study plan FAQHow to use this AI-103 study guide
Microsoft Learn tells you what the exam can assess. This guide adds the practical layer: what the objective usually means, how the Azure service fits into a real AI app, and what type of decision the exam may test.
- Read the domain weights first so you do not over-study smaller sections.
- For each row, open the Microsoft Learn link and skim the official terminology.
- Build at least one small Foundry project that uses a model deployment, an agent, a tool, and Azure AI Search.
- Practise service-selection questions: Foundry Models vs Agent Service vs Azure AI Search vs Content Understanding vs Vision/Language/Speech.
- For RAG and agents, draw the full pipeline rather than memorising isolated terms.
- Before the exam, refresh Microsoft’s official study guide because AI-103 is a newer, fast-moving exam.
Common AI-103 exam traps
Model vs serviceDo not choose an LLM when the requirement is OCR, speech translation, content moderation, or structured document extraction.
RAG is a pipelineGood answers usually include ingestion, extraction, chunking, embeddings, indexing, retrieval, grounding, and evaluation.
Agents need controlsAgent questions often test instructions, tools, memory/state, approvals, least privilege, logging, and guardrails.
Security means keyless firstPrefer managed identity, Microsoft Entra ID, RBAC, private networking, and secretless app configuration where possible.
Responsible AI has stagesLook for prompt shields before generation, content filters during generation, evaluation after generation, and audit traces in production.
Plan and manage an Azure AI solution
| Objective group | Official exam bullet | Topic covered | Main points / likely exam angle | How it appears in Microsoft Learn |
|---|---|---|---|---|
| Choose the appropriate Foundry services for generative AI and agents | Choose an appropriate model for each task, including LLMs, small language models, multimodal models, and Foundry Tools | Model selection in Foundry | Know when to choose GPT-class models, smaller low-latency models, embeddings, multimodal models, image/video models, and tool-based capabilities. Exam questions may describe a workload and ask for the most suitable model family or deployment type. | Learn represents this via Foundry model catalog pages, model capability tables, model cards, regions, deployment types, and benchmark/leaderboard guidance. |
| Choose the appropriate Foundry services for generative AI and agents | Choose the appropriate Foundry services for generative tasks, grounding, vector search, agent workflows, or multimodal processing | Foundry service selection | Map requirements to Foundry Models, Foundry Agent Service, Azure AI Search, Content Understanding, Vision, Speech, Translator, and Language. Expect scenario questions about which service should handle grounding, orchestration, or media processing. | Learn frames Foundry as the unified platform and links out to model, agent, tool, search, and content services. |
| Choose the appropriate Foundry services for generative AI and agents | Choose an appropriate method for retrieval and indexing | Retrieval and indexing strategy | Understand RAG patterns, vector indexes, hybrid search, semantic ranking, chunking, embeddings, and index refresh. Exam may ask whether to use vector search, keyword search, hybrid search, or semantic ranker. | Learn shows classic RAG, hybrid retrieval, vector search, semantic ranking, and how search indexes provide grounding data. |
| Choose the appropriate Foundry services for generative AI and agents | Choose appropriate memory, tool, and knowledge integration services for agent solutions | Agent memory, tools, and knowledge | Know the three agent building blocks: model, instructions, and tools. Match file search, API/function tools, knowledge stores, search, and conversation tracking to agent requirements. | Learn represents agents as models plus instructions plus tools, with search/file/API operations as tool examples. |
| Set up AI solutions in Foundry | Design Azure infrastructure for AI apps and agent-based solutions | Foundry resource and project architecture | Know Foundry resources, projects, endpoints, identity, networking, monitoring, Search, Storage, and app hosting dependencies. Exam may ask about the minimal Azure components for a secure AI app. | Learn describes Foundry as the management plane for projects, model deployments, agents, and AI assets. |
| Set up AI solutions in Foundry | Choose appropriate deployment options | Model deployment options | Understand standard, global/data-zone/regional, serverless API, and managed compute options, plus cost, region, and isolation tradeoffs. Exam may ask which deployment fits latency, data residency, or model-source requirements. | Learn provides deployment option comparisons and notes which model categories support each option. |
| Set up AI solutions in Foundry | Configure model and agent deployments | Deploy and consume models and agents | Know how deployments expose inference endpoints, how an app references deployment names, and how agents are created with prompts/tools. Expect code or portal-step questions. | Learn quickstarts show generating a model response, creating an agent, and running a multi-turn conversation. |
| Set up AI solutions in Foundry | Integrate Foundry projects with CI/CD pipelines | Foundry SDKs and automation | Understand SDK-based provisioning/use, environment variables, endpoints, deployment names, and automation-friendly project configuration. Exam may focus on repeatable deployment and app connection settings. | Learn represents this through SDK and endpoint setup patterns rather than only portal screenshots. |
| Manage, monitor, and secure AI systems | Manage quotas, scaling, rate limits, and cost footprints for model and agent workloads | Capacity, quota, rate limits, and cost | Know how deployment type, tokens, throughput, SKU/capacity, and regional availability affect cost and scale. Exam may ask how to reduce cost or avoid throttling. | Learn’s deployment pages show billing/deployment tradeoffs and supported model/deployment combinations. |
| Manage, monitor, and secure AI systems | Monitor model performance, drift, safety events, and grounding quality | Production monitoring and quality | Track latency, token use, errors, quality scores, safety outputs, groundedness, relevance, and alerts. Expect questions about Application Insights/Monitor and evaluation metrics. | Learn explains observability as evaluation plus monitoring, including dashboards, alerts, token consumption, latency, and quality scores. |
| Manage, monitor, and secure AI systems | Monitor data ingestion quality, search index health, and relevance performance | Search and ingestion monitoring | Know ingestion/index health, freshness, failed documents, relevance, semantic ranker, and hybrid search quality. Exam may ask how to diagnose poor RAG answers. | Learn presents Azure AI Search as the index/query layer for full-text, vector, hybrid, and semantic retrieval. |
| Manage, monitor, and secure AI systems | Configure security, including managed identity, private networking, keyless credentials, and role policies | Identity, network, and RBAC security | Know managed identity, Microsoft Entra ID, role assignments, private endpoints/network isolation, and avoiding embedded keys. Exam may ask for least-privilege or keyless auth design. | Learn positions Foundry and Azure services as Azure resources governed through identity, networking, and access controls. |
| Implement responsible AI across generative AI and agentic systems | Configure safety filters, guardrails, risk detection, and content moderation | Content safety and moderation | Know harmful-content categories, prompt/completion filtering, image/text safety, and moderation decision points. Expect scenario questions about blocking unsafe prompts or outputs. | Learn defines Azure AI Content Safety as detecting harmful user-generated and AI-generated content through APIs and studio tools. |
| Implement responsible AI across generative AI and agentic systems | Apply responsible AI instrumentation, including evaluators, safety evaluations, and explanation tooling | Evaluation and responsible AI instrumentation | Understand built-in/custom evaluators for quality, safety, groundedness, relevance, protected material, and tool-call accuracy. Exam may ask which evaluator to run before deployment. | Learn’s evaluation article shows portal-based evaluations against test data with built-in or custom evaluators. |
| Implement responsible AI across generative AI and agentic systems | Implement auditing through trace logging, provenance metadata, and approval workflows | Tracing, provenance, and auditability | Know why to trace prompts, model responses, tool calls, retrieval sources, and approvals. Exam may ask how to investigate hallucinations or agent mistakes. | Learn’s observability guidance includes tracing, monitoring, and quality/safety signals for generative AI applications. |
| Implement responsible AI across generative AI and agentic systems | Govern agent behavior with oversight modes, constraints, and tool-access controls | Agent governance and controls | Know human-in-the-loop, constrained instructions, least-privilege tool access, approvals, and oversight for semiautonomous workflows. Exam may ask how to prevent risky agent actions. | Learn’s responsible AI overview discusses lifecycle security, governance, controls, checkpoints, and trustworthy agents. |
Implement generative AI and agentic solutions
| Objective group | Official exam bullet | Topic covered | Main points / likely exam angle | How it appears in Microsoft Learn |
|---|---|---|---|---|
| Build generative applications by using Foundry | Deploy and consume LLMs, small models, code models, and multimodal models | Model deployment and inference | Know deployment names, endpoints, SDK/client configuration, token-based billing, and model capability matching. Expect code/config questions around consuming a deployed model. | Learn uses model catalog/deployment documentation and quickstarts to show model selection and inference calls. |
| Build generative applications by using Foundry | Implement retrieval-augmented generation (RAG) in an application | RAG implementation | Know chunking, embeddings, indexes, vector/hybrid retrieval, grounding context, citations, and answer generation. Exam may ask how to reduce hallucinations or improve relevance. | Learn explains RAG as retrieving grounding data from search indexes and passing it to a model for answer generation. |
| Build generative applications by using Foundry | Design workflows, tool-augmented flows, and multistep reasoning pipelines | Tool-augmented workflows | Know how agents/plans call tools, APIs, retrieval, and functions across steps. Exam questions may test which tool schema or workflow pattern fits a scenario. | Learn represents agent workflows as model + instructions + tools, with conversations and tool execution. |
| Build generative applications by using Foundry | Evaluate models and apps, including detecting fabrications, relevance, quality, and safety | Model and app evaluation | Know quality metrics such as coherence/fluency plus RAG metrics such as groundedness and relevance, and safety metrics. Exam may ask what metric identifies hallucination/fabrication. | Learn’s evaluation pages show performance, quality, and safety evaluations before or after deployment. |
| Build generative applications by using Foundry | Integrate generative workflows into applications by using Foundry SDKs and connectors | SDK and connector integration | Know Foundry SDKs/endpoints, OpenAI-compatible APIs, authentication, model deployment names, and app configuration. Exam may include Python snippets. | Learn groups Foundry SDKs and endpoint guidance for Python, C#, JavaScript, and Java development. |
| Build generative applications by using Foundry | Configure an application to connect to a Foundry project | App connection configuration | Know endpoint URL, deployment name, project/resource identifiers, Entra/auth, keys where applicable, and environment variables. Exam may ask which setting is missing in a failing app. | Learn SDK guidance shows how applications connect to Foundry endpoints and projects. |
| Build agents by using Foundry | Define agent roles, goals, conversation-tracking approach, and tool schemas | Agent definition and schemas | Know instructions, goals, constraints, threads/conversations, and tool schemas. Exam may ask what belongs in instructions versus a tool definition. | Learn’s agent overview presents instructions as the agent’s goals/constraints and tools as callable capabilities. |
| Build agents by using Foundry | Build agents that integrate retrieval, function-calling, and conversation memory | Agents with retrieval and functions | Know tool use, search/file retrieval, custom functions/APIs, and preserving state across turns. Expect questions about grounding an agent with private data. | Learn frames tools as the mechanism for search, file operations, and API calls. |
| Build agents by using Foundry | Integrate agent tools, including APIs, knowledge stores, search, content understanding, and custom functions | Agent tool integration | Know when to use API/function tools, Azure AI Search, Content Understanding, and knowledge stores. Exam may ask which tool enables external system actions. | Learn’s agent docs and Content Understanding docs show how tools provide data extraction and actions. |
| Build agents by using Foundry | Implement orchestrated multi-agent solutions | Multi-agent orchestration | Understand role-specific agents, handoffs, shared state, tool isolation, and monitoring. Exam may ask how to divide tasks across agents safely. | Learn’s agent service documentation focuses on hosting/orchestrating AI agents and business-process automation. |
| Build agents by using Foundry | Build autonomous or semiautonomous workflows with safeguards and approval flow controls | Safeguarded autonomy | Know approval gates, oversight, constrained tools, audit trails, and fallback behavior. Exam may ask how to keep a high-risk workflow semiautonomous rather than fully autonomous. | Learn’s responsible AI guidance emphasizes governance, controls, checkpoints, and safe agent lifecycle practices. |
| Build agents by using Foundry | Integrate monitoring into deployed agents, evaluate agent behavior, and perform error analysis | Agent monitoring and error analysis | Track tool-call accuracy, task completion, errors, latency, safety, and trace spans. Exam may ask how to debug an agent that calls the wrong tool. | Learn’s observability article lists agent-specific metrics and tracing support for agent frameworks. |
| Optimize and operationalize generative AI systems | Tune generation behavior, such as prompt engineering and adjusting model parameters | Prompt and parameter tuning | Know temperature, max tokens, system instructions, examples, grounding, and response constraints. Exam may ask how to make output more deterministic or concise. | Learn quickstarts and Foundry playgrounds show iterative testing of prompts, models, and agent behavior. |
| Optimize and operationalize generative AI systems | Implement model reflection, chain-of-thought evaluations, and self-critique loops | Reflection and critique evaluation | Know evaluation patterns where outputs are scored or reviewed by evaluators/models, without exposing private reasoning. Exam may ask how to improve reasoning quality through evaluation loops. | Learn represents this under evaluation/observability using built-in and custom evaluators for quality and reliability. |
| Optimize and operationalize generative AI systems | Set up observability by implementing tracing, token analytics, safety signals, and latency breakdowns | Generative AI observability | Know traces, token usage, latency, error rates, safety events, quality metrics, dashboards, and alerts. Exam may ask which telemetry explains cost or performance spikes. | Learn explicitly describes token consumption, latency, error rates, quality scores, dashboards, alerts, and tracing. |
| Optimize and operationalize generative AI systems | Orchestrate multiple models, flows, or hybrid LLM and rules engines | Multi-model and hybrid orchestration | Know model routing by task, cost/latency tradeoffs, rules for deterministic checks, and fallback models. Exam may ask when to combine rules with an LLM. | Learn model catalog/deployment pages show multiple model families and deployment choices for different task requirements. |
Implement computer vision solutions
| Objective group | Official exam bullet | Topic covered | Main points / likely exam angle | How it appears in Microsoft Learn |
|---|---|---|---|---|
| Design and implement image- and video-generation solutions | Implement a solution that generates images from text prompts and reference media | Image generation | Know compatible model deployments, prompt input, reference media support, base64 output, and saving/using generated images. Exam may ask what model/tool is required. | Learn’s image-generation tool article shows deploying an orchestrator model and image model such as gpt-image-1 in the same project. |
| Design and implement image- and video-generation solutions | Implement a solution that generates videos from text prompts and reference media | Video generation | Know text-to-video and input-media-to-video scenarios, model capability, resolution/duration limits, and preview caveats. Exam may ask which model supports video generation. | Learn’s video generation page describes Sora 2 generating video scenes from text and/or input images or video. |
| Design and implement image- and video-generation solutions | Configure image-editing workflows, including inpainting, mask-based edits, and prompt-driven modifications | Image editing workflows | Understand prompt-driven edits, masks/inpainting, reference images, and edit constraints. Exam may ask which control lets you alter only a region. | Learn represents image editing in Foundry image generation/tool and playground documentation. |
| Design and implement image- and video-generation solutions | Implement workflows to edit generated videos | Video editing workflows | Know generated video lifecycle, reference media, regeneration/editing patterns, and content controls. Exam may focus on tool/model choice and workflow constraints. | Learn represents video generation as a Foundry OpenAI concept with text and media inputs, durations, and resolutions. |
| Design and implement image- and video-generation solutions | Select and apply appropriate generation and editing controls provided by the platform | Generation controls | Know where to adjust prompts, model, size/resolution, duration, reference media, and safety controls. Exam may ask which platform control influences output format or moderation. | Learn’s playground concepts explain testing image/video models and generation/editing experiences in Foundry. |
| Design and implement multimodal understanding workflows | Build a solution that analyzes visual context by using multimodal models | Multimodal understanding | Know models that accept text plus image input and when to use Vision or Content Understanding instead. Exam may ask how to answer questions about an uploaded image. | Learn model catalog lists multimodal model capabilities and Foundry services for visual and text inputs. |
| Design and implement multimodal understanding workflows | Configure apps to produce concise or detailed captions for single or multiple images | Image captions and dense captions | Know Caption versus Dense Captions, supported regions, and when to use Image Analysis 4.0 or alternatives. Exam may ask how to generate descriptions for image assets. | Learn describes captioning and dense captioning as Image Analysis features with region availability notes. |
| Design and implement multimodal understanding workflows | Implement a solution that enables question-answering grounded in visual evidence | Visual question answering/grounded visual answers | Combine image analysis or multimodal models with grounding and evidence. Exam may ask how to make the answer cite or depend on visual content. | Learn represents visual analysis through Image Analysis APIs and multimodal model capability pages. |
| Design and implement multimodal understanding workflows | Configure generation of alt-text and extended image descriptions aligned to accessibility guidelines | Alt text and accessibility descriptions | Know concise alt text versus extended description, captions, and accessibility use cases. Exam may ask what output helps screen-reader users. | Learn has a dedicated Image Analysis use case for auto-generating alt text. |
| Design and implement multimodal understanding workflows | Implement visual understanding by configuring Azure Content Understanding in Foundry Tools to extract visual characteristics | Content Understanding for visual content | Know prebuilt/custom analyzers for images/video, structured outputs, and multimodal extraction. Exam may ask which service extracts visual characteristics into structured data. | Learn describes Content Understanding as analyzing documents, images, audio, and video into structured/searchable data. |
| Design and implement multimodal understanding workflows | Implement video analysis workflows to process and interpret video segments | Video analysis workflows | Know video input, analyzers, segmentation, transcripts/summaries where supported, and structured outputs. Exam may ask how to process media files at scale. | Learn’s Content Understanding quickstart includes document, image, audio, and video analyzer examples. |
| Design and implement multimodal understanding workflows | Configure single-task and pro-mode Content Understanding pipelines | Content Understanding analyzer modes | Know analyzers as reusable configurations combining extraction, AI analysis, and structured output. Exam may ask when to customize an analyzer versus use a prebuilt one. | Learn’s analyzer reference explains analyzers as core building blocks for extraction and output structure. |
| Design and implement multimodal understanding workflows | Implement solutions that identify objects, components, or regions within images or video | Object/region identification | Know tags, objects, bounding boxes, dense captions, people detection, and region-level outputs. Exam may ask which API returns object locations. | Learn’s Image Analysis overview lists object detection, captions, dense captions, tags, people, and OCR features. |
| Implement responsible AI for multimodal content | Implement filters to classify unsafe or disallowed visual content | Visual content safety | Know image moderation, harmful categories, and blocking/flagging unsafe media. Exam may ask which service checks image content safety. | Learn’s Content Safety docs define text/image APIs for detecting harmful material. |
| Implement responsible AI for multimodal content | Detect and mitigate indirect prompt injection by using embedded text in images | Prompt injection in images/OCR | Know prompt shields and the risk of malicious instructions embedded in retrieved documents or images. Exam may ask how to protect a multimodal/RAG app. | Learn describes Prompt Shields as detecting adversarial inputs before generation. |
| Implement responsible AI for multimodal content | Enforce visual policy rules, such as applying watermarks, flagging prohibited symbols, upholding brand usage requirements, and detecting potentially inappropriate content | Visual policy enforcement | Know moderation, post-processing, watermarking policy, brand/prohibited symbol detection workflow, and human review. Exam may ask how to route flagged content. | Learn represents safety/policy enforcement through Content Safety and responsible AI guidance. |
Implement text analysis solutions
| Objective group | Official exam bullet | Topic covered | Main points / likely exam angle | How it appears in Microsoft Learn |
|---|---|---|---|---|
| Apply language model text analysis | Implement solutions to extract entities, topics, summaries, and structured JSON outputs by using generative prompting and Foundry Tools | Text extraction, summaries, JSON | Know NER, key phrases/topics, summarization, custom prompts, JSON schemas/structured outputs. Exam may ask whether to use Azure Language or LLM prompting. | Learn’s Azure Language overview lists prebuilt and custom NLP features; Foundry supports tool-based and LLM-powered patterns. |
| Apply language model text analysis | Configure detection of sentiment, tone, safety issues, and sensitive content | Sentiment, PII, and safety | Know sentiment/opinion mining, PII detection, safety classification, and data limits. Exam may ask how to detect sensitive customer data or negative sentiment. | Learn’s developer guide lists sentiment analysis, PII detection, entity linking, NER, and summarization features. |
| Apply language model text analysis | Build solutions that translate text by using Azure Translator in Foundry Tools or LLM-powered translation flows | Text translation | Know Translator REST APIs/SDKs, supported languages, custom translation, and when an LLM workflow may be needed for context. Exam may ask for production translation at scale. | Learn describes Azure text translation as a cloud REST API for multilingual translation across 100+ languages/dialects. |
| Apply language model text analysis | Customize language model outputs for domain tasks, such as compliance summarization and domain extraction | Domain-specific text extraction | Know custom analyzers, custom NER/classification, prompt constraints, examples, and output schemas. Exam may ask how to extract business-specific fields consistently. | Learn’s Content Understanding custom analyzer tutorial shows creating structured extraction for custom content. |
| Implement speech solutions | Implement workflows to convert speech to text and text to speech for agentic interactions | Speech to text and text to speech | Know STT, TTS, real-time/batch transcription, neural voices, and voice-enabled agent interactions. Exam may ask which feature handles live voice input or response audio. | Learn’s Speech overview covers transcribing speech, producing natural-sounding voices, and live AI voice conversations. |
| Implement speech solutions | Integrate speech as an agent modality, including custom speech models | Custom speech for agents | Know when to customize recognition for domain vocabulary, accents, or noisy environments. Exam may ask how to improve transcription accuracy for specialized terms. | Learn’s custom speech guide explains fine-tuning speech recognition for real-time STT, speech translation, and batch transcription. |
| Implement speech solutions | Enable multimodal reasoning from audio inputs | Audio input reasoning | Convert audio to text, preserve speaker/context metadata where needed, then use an LLM/agent for reasoning. Exam may ask what must happen before a text-only model reasons over audio. | Learn represents speech-to-text as the conversion layer for real-time and batch transcription. |
| Implement speech solutions | Translate speech into other languages by using language models and Foundry Tools | Speech translation | Know real-time speech-to-speech and speech-to-text translation, supported languages, and integration with Translator/LLMs. Exam may ask for live multilingual meeting translation. | Learn’s Speech Translation overview describes real-time, multi-language speech-to-speech and speech-to-text translation. |
Implement information extraction solutions
| Objective group | Official exam bullet | Topic covered | Main points / likely exam angle | How it appears in Microsoft Learn |
|---|---|---|---|---|
| Build retrieval and grounding pipelines | Ingest and index content, such as documents, images, audio, and video | Multimodal ingestion and indexing | Know ingestion sources, media processing, analyzers, chunking, embeddings, and search indexing. Exam may ask how to prepare mixed media for RAG. | Learn’s Content Understanding docs emphasize transforming documents, images, audio, and video into structured, searchable data. |
| Build retrieval and grounding pipelines | Configure semantic search, hybrid search, and vector search for grounding | Semantic, hybrid, and vector grounding | Know BM25/full-text, vector similarity, semantic ranker, reciprocal rank fusion, and when hybrid improves relevance. Exam may ask which search mode balances precision and recall. | Learn’s hybrid search overview explains parallel full-text and vector queries merged with RRF; semantic ranker reranks by meaning. |
| Build retrieval and grounding pipelines | Implement enrichment by using custom or built-in skills for text, images, and layout | Enrichment skills and indexing | Know built-in/custom skills, OCR, image/layout extraction, chunk enrichment, and index fields. Exam may ask how to enrich raw files before retrieval. | Learn represents Azure AI Search as the pipeline/index layer that supports full-text, vector, multimodal, enrichment, and secure search scenarios. |
| Build retrieval and grounding pipelines | Configure RAG ingestion flow, including documents and using optical character recognition (OCR) | RAG ingestion and OCR | Know document cracking, OCR, chunking, embeddings, metadata, index design, and refresh. Exam may ask how to make scanned PDFs searchable for RAG. | Learn’s RAG and Search docs show indexing and retrieval patterns for grounding model responses. |
| Build retrieval and grounding pipelines | Connect retrieval pipelines directly to workflows and agent tools | Retrieval as an agent tool | Know how agents call search/knowledge tools, retrieve context, and use it inside a workflow. Exam may ask how to ground an agent in enterprise content. | Learn’s agent overview describes tools as access to data/actions, including search and file operations. |
| Extract content from documents | Extract information by using multimodal pipelines that combine OCR, layout analysis, and field extraction | OCR, layout, and field extraction | Know prebuilt analyzers such as invoices, custom schema extraction, and multimodal pipeline outputs. Exam may ask which service extracts fields from scanned forms. | Learn’s Content Understanding quickstart shows invoice/document analysis and structured extraction from documents, images, audio, and video. |
| Extract content from documents | Produce clean, grounded representations to use with agents and RAG by using Content Understanding | Clean grounded representations | Know converting unstructured media into structured/Markdown/searchable representations for downstream LLMs and agents. Exam may ask how to prepare documents for reliable RAG. | Learn describes Content Understanding as transforming unstructured content into structured, organized, searchable data. |
| Extract content from documents | Implement analyzers for generating structured or markdown outputs for downstream reasoning by using Content Understanding | Analyzers and markdown/structured outputs | Know analyzers, field schema, prebuilt versus custom, and outputs for downstream reasoning. Exam may ask how to guarantee consistent JSON/Markdown from varied documents. | Learn’s analyzer reference defines analyzers as reusable configurations for extraction, AI-powered analysis, and structured data output. |
Fast AI-103 exam revision tips
- Think in scenarios. Microsoft role-based exams usually describe a business requirement and ask which Azure service, model, deployment, security option, or monitoring approach best fits.
- Know the boundaries. Be able to separate Foundry Models, Foundry Agent Service, Azure AI Search, Content Understanding, Vision, Speech, Translator, Language, and Content Safety.
- For RAG questions, remember the pipeline: ingest → extract/OCR → chunk → embed → index → retrieve with vector/hybrid/semantic search → ground the prompt → evaluate groundedness/relevance.
- For agents, remember the pattern: model + instructions + tools + memory/conversation state + monitoring + guardrails.
- For security questions, prefer least privilege: managed identity, Microsoft Entra ID, RBAC, private networking where required, and no hard-coded keys unless the scenario forces key-based auth.
- For responsible AI questions, look for the control point: pre-generation prompt shield, generation content filter, post-generation evaluation, trace/audit, or human approval.
Suggested AI-103 study order
Step 1Review Microsoft Foundry concepts: projects, models, deployments, endpoints, agents, tools, and evaluations.
Step 2Build a small generative AI app with a model deployment, prompt settings, SDK authentication, and basic monitoring.
Step 3Add RAG with Azure AI Search: ingestion, chunking, embeddings, vector/hybrid search, semantic ranking, and citations.
Step 4Build an agent with instructions, tools, file/search grounding, function/API calls, conversation state, and safety controls.
Step 5Review multimodal services: Vision, Speech, Translator, Language, Content Safety, and Content Understanding.
Practical revision: For every objective, ask: which Azure service handles this, what input/output does it expect, how is it secured, how is it monitored, and how would I evaluate whether it works?
AI-103 exam FAQ
Is AI-103 replacing AI-102?
AI-103 is a separate Azure AI Apps and Agents Developer certification path focused on Microsoft Foundry, generative AI apps, agentic solutions, RAG, multimodal AI, and production AI patterns. AI-102 remains a different Azure AI Engineer exam track unless Microsoft changes the credential structure again.
Is Microsoft Learn enough to prepare for AI-103?
Microsoft Learn is the primary source for the exam scope, but you should also build hands-on examples. The exam can test applied decisions: which service to choose, how to configure retrieval, how to secure a deployment, how to monitor a model, and how to evaluate agent behaviour.
What should I study first for AI-103?
Start with Microsoft Foundry basics and generative AI app development. Then study RAG with Azure AI Search, agents and tools, responsible AI, observability, and finally the smaller multimodal sections: vision, speech, text analysis, and content extraction.
What is the hardest part of AI-103?
The hardest part is usually connecting services together. Many questions are not “what is Azure AI Search?” but “how do you use search, embeddings, tools, monitoring, and safety controls together in a production AI app or agent?”
How often should I refresh this guide?
Refresh it whenever Microsoft changes the AI-103 study guide or Foundry documentation. AI and agent services are moving quickly, so stale terminology can become a ranking and exam-preparation problem.
