Draw the boundary first
Not every step suits a model. We only hand over steps with a clear input and a checkable output — turning a client email into structured tickets, for example. Anything involving money, commitments or legal wording stays with a person.
Define acceptance next
Every integration point needs a measurable condition: not "it looks smarter", but "field completeness went from X to Y" or "manual rework dropped by N". If it misses, the workflow falls back to manual instead of another round of prompt tuning.
Connect the model last
Once the first two steps are done, wiring the model is usually quick: a replaceable interface, a version log for prompts, and a fallback path. Most of the time goes into making the business rules explicit.
Where we are now
Agents are best at repetitive, structured, re-readable work, and worst at judgement with liability attached. Treat one like an intern, not like an owner.