What Enterprise AI Integration Actually Involves.
Production effort sits in identity, approved data, system contracts, failure handling, observability and ownership - not only in the model or prompt.
The boundary is the design
Before anything is built, agree what the workflow may read, what it may write, and what it must never touch. Those three lists are the actual architecture; everything else is implementation detail.
Identity and permissions come before features
An automated workflow acts on behalf of someone. Decide whose authority it uses, how that is recorded and how access is revoked, or you will retrofit it under audit pressure later.
Plan for the states nobody demonstrates
Timeouts, partial writes, duplicate submissions and stale data are the normal operating conditions of an integrated system. A workflow that has no defined behaviour for them will invent one.
Observability is a requirement, not a phase
You cannot govern what you cannot see. Logging of inputs, decisions, actions and outcomes has to be part of the first release, because it is what makes every later improvement arguable.
Write down what the workflow may read, write and never touch. Then settle identity, failure states and observability before adding capability.