Azure AI Foundry: The Honest Enterprise Review

I deployed a bank compliance system on Azure AI Foundry — real documents, real auditors, real SLAs. This is the review I wanted to read before we started: what's genuinely good, what's marketing, and what will bite you in week three.

The governance story is the product. If your buyer is a bank's risk committee, Foundry's pitch lands: models, deployments, content filters, and evaluations under one roof, wired into the Azure identity and compliance machinery the bank already trusts — Entra ID, private endpoints, regional data residency. We cleared infosec review in weeks, not quarters, largely because nothing about the platform was new to the security team. The model catalog breadth is real — OpenAI, Phi, and a long tail of open-weight models behind one deployment pattern, which makes the route-by-difficulty cost strategy practical. And prompt flow evaluations, while not magical, put eval runs where the auditors can see them — that visibility bought us credibility no notebook ever could.

The "build an agent in minutes" demos are true and useless — the minutes-agent is the demo-grade artifact this entire blog warns about. The agent tooling is young: orchestration beyond simple patterns still wants LangGraph or your own state machine, with Foundry as the model and governance layer underneath. Treat "end-to-end agent platform" claims the way you treat all platform claims: the last 30% is still your engineering.

Quota choreography: capacity for top models varies by region and moves; design multi-region failover from day one or discover it during your launch week. Cost opacity by default: turn on detailed token attribution immediately — the gap between list price and effective cost (retries, filters, evals) surprised our finance partner by ~25%. Content-filter friction: compliance text trips false positives; budget tuning cycles for filter policies, and document every exception — that paper trail is gold in audit. Version pinning: auto-upgrading model versions is convenient until an upgrade shifts behaviour under your evals; pin, test, then move.

Bottom line: Foundry won't make your architecture decisions — but it makes the right architecture deployable inside a regulated enterprise, and in banking that's most of the war.