Start with the chores
Look at a normal week and find the jobs nobody would miss doing: the first draft of a reply, minutes of a meeting, a product description, a tidy-up of a messy spreadsheet. These are safe places to begin, because a rough answer is still useful and a wrong one is easy to spot. An AI tool is quick at producing something reasonable from a clear request. It has no idea whether that something is true for your business, so the person who asked still reads it before it is used.
The checking pile
The second pile is work AI can speed up, where a mistake would cost you. Prices, dates, stock, legal wording, anything sent to a customer. Here the tool does the legwork and a named person checks the result. Write down who that person is. In my experience the problems come from the unclear cases, where everyone assumed somebody else had read it.
What stays with people
Some work is the reason customers come to you. Deciding whether to give someone a second chance. Telling a customer something they will not enjoy hearing. Judging a risk. A tool can lay out options, and it can be a useful sounding board, but the decision and the responsibility are yours. Customers can tell the difference, and they mind.
A fair test for any new job
Before handing a task to AI, ask three things. Would a wrong answer be obvious? Would it matter much if one slipped through? Is there private information in what I am about to type? Obvious, minor and no: go ahead. Anything else belongs in the checking pile, or stays with you.
Where to go from here
Pick three chores, give each to one person, and try for a fortnight. Then look at what happened. That small, slightly boring exercise teaches a business more than any amount of reading, and it is what the last half hour of my workshop is for.