Domain 1: Plan and manage an Azure AI solution
- Choose the right Foundry services and models for a scenario: pick from the model catalog (Azure OpenAI and Azure AI model deployments), and match capabilities (chat, reasoning, embeddings, multimodal, vision, speech) plus cost, latency, and region to the requirement.
- Set up a Microsoft Foundry (Azure AI Foundry) project and hub, then create model deployments with a chosen model, version, and capacity (tokens-per-minute quota); the deployment provides the endpoint your app calls.
- Prefer keyless, identity-based access: use managed identity and Microsoft Entra ID with role-based access control (for example the Azure AI User role) instead of API keys, and store any required secrets in Azure Key Vault.
- Lock down networking with private endpoints and disabling public network access so traffic to Foundry, Azure AI Search, and Azure OpenAI stays on your virtual network.
- Build CI/CD for AI apps: source-control prompts, flows, and infrastructure-as-code, then automate deployment of model configurations and agent definitions through pipelines with promotion across environments.
- Manage cost and quota by selecting standard pay-as-you-go versus Provisioned Throughput Units (PTUs) for reserved capacity, monitoring tokens-per-minute limits, and handling HTTP 429 throttling with retry and exponential backoff.
- Monitor solutions with Azure Monitor and diagnostic settings to Log Analytics, capturing token usage, latency, errors, and traces; Foundry observability adds tracing for prompts, tool calls, and agent runs.
- Apply responsible AI: enable Azure AI Content Safety filters on prompts and completions, run Foundry evaluators (groundedness, relevance, coherence, fluency, safety) against test datasets, and use groundedness detection to catch ungrounded or hallucinated output.
Domain 2: Implement generative AI and agentic solutions
- Build generative apps by deploying a model in Foundry and consuming it from Python via the Azure OpenAI SDK or the Azure AI Inference client, sending chat messages and reading completions.
- Apply prompt engineering: craft a clear system message for persona and output format, use few-shot examples, and tune parameters such as temperature, top_p, and max tokens (low temperature for deterministic, grounded answers).
- Implement Retrieval-Augmented Generation (RAG): retrieve relevant chunks from Azure AI Search (vector, hybrid, or semantic) and inject them as grounding context so the model answers from your data without retraining.
- Deploy and consume models from the catalog including Azure OpenAI models and open/partner models via Azure AI model deployments, choosing serverless or managed compute as needed.
- Build agents with the Foundry Agent Service: define an agent with a model, role/instructions, and tools; create conversation threads that persist message history and memory so each user session stays isolated.
- Give agents tools - built-in tools such as file search over a vector store, code interpreter, and Azure AI Search grounding, plus custom function calling where the model returns a function name and JSON arguments for your app to execute.
- Orchestrate multi-agent solutions with patterns like a supervisor/orchestrator delegating to specialist agents, sequential handoff, and group chat; connected agents let one agent call another as a tool.
- Evaluate and improve generative and agent solutions using Foundry evaluators and test datasets (groundedness, relevance, task success), and iterate on prompts, retrieval, and tool definitions.
- Add observability with tracing to inspect each run's prompts, retrieved context, tool calls, and token usage, enabling debugging and regression testing of agents in production.
Domain 3: Implement computer vision solutions
- Generate and edit images using image-generation models deployed in Foundry (for example DALL-E-family and other catalog image models), controlling output with prompts and edit/inpainting requests where supported.
- Use multimodal models to understand images and video: send an image alongside a text prompt to a vision-capable model to caption, answer questions about, or reason over visual content.
- Apply Azure AI Content Understanding to extract structured insights from multimodal content (documents, images, audio, and video) with a schema-driven, analyzer-based approach.
- Use Azure AI Vision for prebuilt image analysis features such as captions, tags, object detection, and Read OCR when a task-specific API fits better than a general multimodal model.
- Enforce responsible AI for multimodal scenarios: run image and video content through Azure AI Content Safety (including image moderation across hate, sexual, violence, and self-harm categories) and filter generated images for harmful or protected content.
- Choose the right approach per scenario: general multimodal model for open-ended visual reasoning and generation, Azure AI Vision for prebuilt vision tasks, and Content Understanding for structured extraction from mixed media at scale.
Domain 4: Implement text analysis solutions
- Use Azure AI Language for text analysis: named entity recognition, key phrase extraction, language detection, and PII detection and redaction, returning results with confidence scores.
- Run sentiment analysis (positive/negative/neutral with scores) and opinion mining to link sentiment to specific aspects, plus extractive and abstractive text summarization.
- Produce structured output from unstructured text by prompting a model to return JSON, or by combining Language features to populate a defined schema for downstream systems.
- Translate text across languages with Azure AI Translator, supporting one source to multiple targets in a single request and custom terminology where needed.
- Implement speech scenarios with Azure AI Speech: real-time and batch speech-to-text (STT), and text-to-speech (TTS) with neural voices controllable via SSML for rate, pitch, and emphasis.
- Select between a task-specific Language/Speech API and a general model: prebuilt services give deterministic, cost-effective results for standard tasks, while models offer flexible, prompt-driven extraction and generation.
Domain 5: Implement information extraction solutions
- Build retrieval and grounding pipelines in Azure AI Search: create an index, connect a data source, and use an indexer with a skillset to crack documents, apply OCR and enrichment, and populate searchable fields.
- Enable vectorization and integrated vectorization so Azure AI Search automatically chunks and embeds content during indexing, producing a vector index for RAG without a separate embedding pipeline.
- Support vector, hybrid, and semantic retrieval - use an embedding model to vectorize queries, combine keyword and vector results (hybrid), and apply semantic ranking to reorder results for grounding.
- Extract fields from documents with Azure AI Document Intelligence using prebuilt models (invoice, receipt, ID, layout) or custom models, returning fields, tables, key-value pairs, confidence scores, and bounding regions.
- Use Azure AI Content Understanding for schema-driven extraction across documents, images, audio, and video when you need structured output from mixed or multimodal sources.
- Feed extracted and enriched content into the grounding layer so RAG apps and Foundry agents can retrieve accurate, source-grounded context at run time.
AI-103 exam tips
- Think Foundry-first: nearly every scenario is framed inside a Microsoft Foundry (Azure AI Foundry) project - know how projects, hubs, model deployments, connections, and the Foundry Agent Service fit together.
- Weight your study by domain: generative AI and agents (Domain 2) plus plan/manage (Domain 1) together dominate the exam, so master model deployment, RAG, agent tools/threads/memory, function calling, evaluators, and keyless security first.
- Practice in Python: the exam is developed with Python, so be comfortable with the Azure OpenAI SDK, the Azure AI Inference client, the Foundry Agent Service SDK, and Azure AI Search queries in code.
- Drill service selection: build a decision table mapping needs (ground answers in your data, extract form fields, moderate images, transcribe speech, understand multimodal content) to the exact Foundry service.
- Know the security story cold: managed identity and Entra ID over keys, keyless access, private endpoints with public access disabled, Key Vault for secrets, and Content Safety plus evaluators for responsible AI.
- Memorize the RAG and retrieval details: vector vs hybrid vs semantic search in Azure AI Search, integrated vectorization, embedding models, and how retrieved context grounds a model or agent.
- Understand agent internals: the difference between a thread (per-user conversation and memory), a tool (built-in like file search/code interpreter or a custom function), and multi-agent orchestration patterns such as supervisor and connected agents.
- Read each scenario for the deciding constraint - cost, latency, network isolation, no-training-allowed, groundedness, or on-premises - because that constraint usually eliminates all but one answer.
Study guide FAQ
How is AI-103 structured and scored?
AI-103 is an associate-level Microsoft exam delivered in English, lasting 120 minutes, with a passing score of 700 out of 1000. Microsoft does not publish an exact question count, but expect roughly 40-60 scenario-based questions. Scoring is scaled and weighted rather than a raw percentage, and there is no penalty for wrong answers, so answer every question. The exam is developed with Python.
Is CertGrid's AI-103 practice official Microsoft material?
No. CertGrid is an independent practice platform and is not affiliated with or endorsed by Microsoft. Our questions are original and written to mirror the official AI-103 skills outline and difficulty, so you can rehearse the exam objectives and identify gaps before test day. Always cross-reference the current Microsoft Learn exam page for the authoritative skills measured.
What prerequisites or experience do I need for AI-103?
AI-103 assumes hands-on application-development experience in Python and familiarity with generative AI concepts (prompts, models, RAG, agents) and core Azure services. You should be comfortable deploying and consuming models in Microsoft Foundry, calling Azure AI services from code, and working with Azure AI Search. There is no mandatory prerequisite exam, but prior Azure and AI fundamentals knowledge helps.
How should I focus my study across the five domains?
Prioritize generative AI and agentic solutions (Domain 2) and plan/manage (Domain 1), which together carry the most weight - roughly 30-35% and 25-30%. Master model deployment, RAG, Foundry agents (roles, tools, threads/memory, function calling), multi-agent orchestration, evaluators, and keyless security. Then round out with computer vision, text analysis, and information extraction (Domains 3-5), each around 10-15%, focusing on service selection and Content Understanding, Document Intelligence, and Azure AI Search grounding.
How does AI-103 relate to the retired AI-102?
AI-103 is the active replacement for AI-102 (Azure AI Engineer Associate), which has been retired. Compared with AI-102, AI-103 is Foundry-first and shifts emphasis toward generative AI, RAG, and agents built on the Foundry Agent Service, uses current terminology (Microsoft Foundry, Azure AI model deployments, Content Understanding), and is developed with Python. If you studied AI-102, expect much deeper coverage of agents, multi-agent orchestration, evaluation, and observability, with less standalone service trivia.