Where AI should stop.

Start with the repeat.

A team may receive the same kind of enquiry every day, search the same documents or prepare the same brief. Those are useful places to begin because the work can be described and checked.

“Add AI” is not a brief. Name the input, the output and the person who needs it.

Define what it may know.

A system should get the context needed for its task, not every file the business owns. The access boundary is part of the design, not an implementation detail to settle later.

Preparation can be assisted. Accountability cannot be outsourced.

Make review visible.

An assistant can structure an enquiry and draft a response. A person should approve a promise to a customer, a price or a consequential change. The interface should make that transition obvious.

Improve from real use.

Look at where the system saves time and where it creates rework. Change the workflow around those observations. The goal is a better operation, not a more impressive demo.