If a stakeholder told me, “We need AI in the business,” I’d ask to see a normal working day. Show me the replies people write over and over. Show me the weekly report that takes an hour to pull together. Somewhere in that routine there’s probably a task small enough to test and familiar enough that people can tell when the output is off.
Customer support is a useful example. A published study of 5,172 support agents found that access to an AI assistant increased issues resolved per hour by 15% on average, with results varying across workers. I’d take that as a reason to run a focused pilot, using the team’s own workload and measures.
Sit beside the person who’ll use it
Say the assistant drafts replies from approved policies. Put that draft where the agent already reads the ticket. Let them see which policy it used, change the wording, and discard the answer if it doesn’t fit. Ask them which edits they keep having to make. Google’s People + AI Guidebook covers feedback, control, and recovery, all of which become real design decisions on that screen.
I’d start with one queue and compare speed, accuracy, reopen rates, and customer satisfaction. If the team is replying faster but customers keep coming back with the same issue, I’d look at the replies with the agents. They’ll usually be able to tell you where the tool guessed or missed an important detail.
