jcardena.com › Blog › AI Foundry by Contract: Delivering Governed AI on Azure
184 posts
EN ES

AI Foundry by Contract: Delivering Governed AI on Azure

A free field guide to Microsoft Foundry on a governed Azure landing zone: projects, roles, models and quota, private networking, safety, evaluations, tracing, the AI gateway and agents by contract.

TL;DR: A model deployment takes minutes. Knowing who may call it, what it may see, what it costs and whether its answers are right takes design. AI Foundry by Contract is a free field guide that builds Microsoft Foundry on a governed landing zone, so every project, deployment and agent has an owner, every consumer has a limit, and supported calls leave a log. It is Book 2 of the GDCD on Azure series.

Ungoverned AI adoption often begins with one deployment and one shared credential, created outside the landing zone. It works. Then the questions arrive. Who else has the key? Which data can the model see? Why did the bill triple? Is the answer correct, and can anyone prove it?

This guide treats those questions as design inputs. It applies Governance-Driven Contract Design (GDCD) to AI on Azure: start from the business outcome, turn it into guarantees, and write each guarantee as a contract at the door where it can be enforced. The model proposes. The contract decides.

A lit stone archway that every visitor passes through, like an AI gateway

What is in the book

  • Where Foundry lives in the landing zone. Foundry resources and projects placed in the management group tree and subscriptions from Book 1, so AI inherits the same guardrails as every other workload.
  • Roles for AI work. The Foundry roles, who gets which, and why keys give way to Microsoft Entra ID.
  • Models, deployments and quota. Approved models through supported Azure Policy controls and deployment pipelines, deployment types and data residency, and quota as a contract, not a surprise.
  • Private by default for AI. Private endpoints, DNS and network isolation for AI services.
  • Safety and evaluations. Content safety as a guardrail, and evaluations as repeatable evidence of measured quality and safety before and after release.
  • Tracing and visibility. Instrumented workloads that record calls, tool activity, latency and errors, so a reviewer can see what was called and what came back, with sensitive data protected.
  • Tokens are spend: the AI gateway. Azure API Management as the gateway to supported model endpoints, with identity and network rules that block bypass, with token limits, token metrics and, where configured, semantic caching per consumer.
  • Agents by contract. Approved tools, identity per agent and human approval at the doors that matter.
  • Templates and regulatory mapping. An appendix with contract templates and a map to common regulatory expectations.

Why now

AI on Azure has moved from pilots to production. Spend, data exposure and answer quality now need owners and controls. The guide treats each one as a guarantee with an owner and a control.

It builds on Governing Azure by Contract, which sets up the landing zone, roles, policy and cost controls this book assumes. Read that one first if your platform is not yet governed.

How it was checked

Product names, roles, policy effects and gateway features were checked against Microsoft documentation in November 2026. The text then went through four rounds of AI-assisted review, with several AI models prompted to read as an Azure architect, a security and FinOps buyer and a technical editor. These were review tools, not an independent audit. Each factual finding they raised was checked against Microsoft documentation before a change was accepted, and findings that did not hold up were rejected.

Read it

The guide is free to read online or download as a PDF. The Foundry model catalog lives on the Azure hub.

A useful first question for any AI workload on Azure: for every model you call today, can you name its owner, its limit and where its log is? The model proposes. The contract decides.

JC
Juan Cardena
Enterprise AI & Agentic Systems Architect

Enterprise architect with 25 years across web, software, data, and AI. MIT CDAO ’25. Writing on agentic AI in production.