What's changed: Added per-section figures (cert-figure-retrofit). New AI-103 Chapter 1 (plan & manage: project/resources, model selection = capability/modality/context/cost/regional availability/routing, deployment = serverless vs managed compute, TPM quota/cost, security = managed identity/keyless/RBAC/private endpoint, monitoring/evaluation/Content Safety/Responsible AI)
1.1Planning AI Solutions and Provisioning Resources
Plan how to build AI apps/agents on Microsoft Foundry from a developer’s view—organizing projects and resources, selecting a fitting model (capability, modality, context length, cost, regional availability), deployment options, and estimating quota (TPM) and cost.
AI-103 certifies a developer who designs, implements, and operates production-grade generative AI apps and agents on Microsoft Foundry, building on the conceptual grounding of AI-900/AI-901. The first pillar is "plan and manage"—deciding which resources, which models, which regions, and what capacity and cost to provision. Getting this wrong causes rework and cost overruns later.
1.1.1Organizing projects and resources
Foundry work is organized into projects. A project gathers deployed models, connected Azure AI services (Language, Speech, Vision, Content Understanding, etc.), data, agents, and evaluation results, and is backed by Azure resources (subscription/resource group). Developers separate projects per environment (dev/staging/prod) and bind services and data stores via connections, making credentials and access easier to manage centrally.
1.1.2Selecting a model
The model catalog (Foundry Models) offers many options—Microsoft, OpenAI, open models. For production design, beyond AI-901’s capability, modality, context length, and cost, a developer also weighs regional availability (is the model in your target region?), latency/throughput, and data handling (regulatory/residency). Keep "smallest sufficient model" as the rule, but design routing—large models for hard tasks, small models for routine ones.
1.1.3Deployment options, quota, and cost
You deploy a model and use it via an endpoint + auth. Deployment options include the easy pay-as-you-go serverless API type and the managed compute type that assigns dedicated compute—consider the latter when you need steady throughput or isolation. Generative AI cost is billed by tokens (input + output), and throughput is capped by quota such as TPM (tokens-per-minute). At design time, estimate TPM from expected traffic and request quota increases as needed.
| Aspect | Serverless API | Managed compute |
|---|---|---|
| Billing | Pay-as-you-go (per token) | Allocated compute time |
| Best for | Variable load, quick start | Steady throughput, isolation |
| Throughput | Capped by TPM quota | Depends on allocated capacity |
Common: (1) "is the model in the target region" → check regional availability. (2) Generative AI throughput cap = TPM (tokens-per-minute) quota. (3) Cost = input + output tokens. (4) Production needing steady throughput/isolation = managed compute; easy pay-as-you-go = serverless API. (5) Routing design: "smallest sufficient model + large only for hard tasks."
Watch out: (1) quota (caps like TPM) and cost (token billing) are different—raising quota does not change that you pay for what you use. (2) Regional availability varies per model—features are not in every region. (3) Deployment type (serverless/managed) is chosen by requirements; neither is always correct.
1.1.4Section summary
- Organize by project, split per environment, bind services/data via connections
- Model selection: capability/modality/context/cost + regional availability/latency/data handling; design routing
- Deployment: serverless API (pay-as-you-go) vs managed compute (steady/isolated)
- Throughput capped by TPM quota, cost by input + output tokens; estimate and request TPM
Sign in to track progress — Log in.
Quick check
(just a quick review)Q1. Which quota represents the throughput cap of a generative AI model?
Q2. Which deployment option suits production needing steady throughput and workload isolation?
Q3. When choosing a model, what is the concern of checking whether it is offered in your target region?
Q4. Which correctly describes the relationship between quota and cost?
Q5. What is the common unit for separating dev/staging/production in Foundry?
Q6. Which is an appropriate production model-operation design to control cost/latency while meeting quality?