Every conversation about AI systems starts at the wrong altitude. The question arrives as "how do we transform the department" when the work that actually hurts is an evening, twice a month, for two people who are good at their jobs but should not be spending them this way.
The scale trap
Big scopes fail in a specific way: they cannot be validated. If a system is meant to change how twenty people work, its acceptance test becomes a project of its own, and the honest answer to "does this work?" arrives after the budget does.
Small scopes have the opposite property. One workflow, one decision, one signature — you can put the output in front of the person who owns it and get a real answer in a week.
What small still requires
Small does not mean shallow. A tender desk that reads a requirement list, sources every answer from the firm's own material and names its gaps is doing real engineering: intake, extraction, retrieval, assembly, and a validation rule strict enough to refuse.
The layers do not shrink. Only the number of people the system touches does.
A test before you build anything
Name the person who signs the output, and name the moment in their week that gets shorter. If either answer is vague, the scope is too big or the problem is not yet understood — and no architecture will repair that.