AI-103 · Planning & Managing Solutions · Module 1 of 16 · Free preview

Choosing the Right Foundry Services

Picks up right after AI-901 · 13 min read
What this module covers

AI-901 taught you to recognize which workload a scenario needs. AI-103 is where you actually pick, provision, and code against the right Foundry service to build it for real.

"I know it's agentic AI. Now what?"

Aditi's AI-901 certification confirmed she could correctly identify that the return-approval system needed an agentic solution, and her earlier work had built and tested it manually in the Foundry portal. Now her manager wanted it built as a real, deployable application, with code, not something demonstrated by clicking through a portal each time. That gap, from correctly recognizing a workload to actually building a production application around it, is exactly what AI-103 tests.

The Foundry SDK: the portal's capabilities, in code

Everything explored manually in the Foundry Playground during AI-901 has a corresponding SDK call: deploying a model, sending a prompt, configuring an agent, all callable from Python rather than clicked through a UI. AI-103 assumes you can read this code, `azure-ai-projects` and related packages, understand what a given snippet does, and reason about how to modify it, not necessarily write a large application from a blank file under exam pressure.

Choosing services for a real requirement

A real application rarely needs just one Foundry capability in isolation. Aditi's actual build needed: a language model for reasoning about the return policy, an agent wrapping it with the order-lookup tool, and, once the finance team's invoice-extraction request came in later, Content Understanding layered into the same project. Planning a solution means mapping each part of the requirement to the specific service that handles it, then deciding how they connect.

Planning for cost and the solution lifecycle

Beyond picking the right service, planning includes realistic capacity sizing (Module 8 of AI-901's course covered the tradeoff of over- versus under-provisioning), and thinking about the solution's lifecycle, how a deployed model gets updated, retired, or replaced as better options become available, rather than treating a deployment as a one-time decision that never needs revisiting.

Check your understanding

Five questions, all from this module. Get 4 out of 5 to mark it complete.

1. What was the actual gap between Aditi's AI-901 certification and what her manager now needed?

2. What level of Python skill does AI-103 assume, according to this module?

3. What does the Foundry SDK let a developer do?

4. Why did Aditi's real build need more than one Foundry capability?

5. What does planning for the solution lifecycle involve?