AI-103 Tests Whether You Can Build, Ground and Govern AI Apps and Agents on Microsoft Foundry
AI-103, the successor to the retired AI-102, rewards choosing the right Foundry model and tools, grounding answers with retrieval, and securing, evaluating and controlling agents in production.
Check Your Readiness →Most candidates understand Microsoft Certified: Azure AI Apps and Agents Developer Associate (AI-103) concepts — and still fail. This exam tests how you apply knowledge under pressure.
The steps follow the AI-103 domains, led by generative AI and agentic solutions (30-35%) and planning and managing the solution (25-30%). Most scenario items describe a business need and a constraint, and the right answer meets both with a supported Foundry capability.
Give the agent a broadly scoped identity and let it call the refund API whenever its reasoning decides a refund is due.
Expose the API as a narrowly defined tool with a clear schema, run it under a least-privilege identity and require human approval for high-impact actions. The outline explicitly tests safeguards and approval flows for autonomous workflows.
Fine-tune the model on old policy PDFs or raise the temperature so it gives fuller answers.
Ground the app with RAG over an index of the current policies using hybrid search and semantic ranking, instruct it to answer only from retrieved content, and track groundedness with evaluators. Lower temperature also reduces variation.
Keep public network access on and restrict it with IP allow lists while using account keys from the app.
Use private endpoints for the Foundry and search resources, disable public network access, and authenticate with managed identity. This meets the network requirement and removes secrets at the same time.
Microsoft's guidance for Foundry and Azure AI services is keyless authentication with Microsoft Entra ID, typically a managed identity used through DefaultAzureCredential, with RBAC roles scoped to the resource. Answers that store keys in code or config files are usually distractors when a managed identity option is offered.
Fine-tuning changes a model's style, format or task behaviour; it is not the right way to give it current or proprietary facts. When answers must reflect company documents that change, the expected design is retrieval-augmented generation over an index.
In Azure AI Search, hybrid search runs keyword and vector queries together and merges the results with Reciprocal Rank Fusion. Semantic ranking is a separate re-ranking step applied to the top results, so it complements hybrid search rather than replacing it.
Content filtering in Foundry evaluates both the prompt and the completion across harm categories. Prompt Shields add detection of direct jailbreak attempts and indirect prompt injection hidden in documents, emails or images the model reads.
Model deployments have quotas measured in tokens per minute with derived request limits, and exceeding them returns HTTP 429. The fix is retry with backoff, a higher quota, spreading load across deployments or provisioned throughput, not rewriting the prompt.
Strengthen weak areas with exam-style practice questions and detailed explanations.
Simulate the real exam experience and assess your readiness under timed conditions.
Review key concepts, objectives, and exam topics in one place.
Get personalized explanations, learning recommendations, and instant answers.
Follow a structured learning path designed to help you prepare efficiently.
Find out whether agents, grounding or governance is costing you the most AI-103 points before you book.