The request is usually for a tool. The problem is usually the work around it.
01 · The question as it arrives
“We should be using AI.” Sometimes it arrives from the board, sometimes from a competitor’s press release, and sometimes from a team that is simply tired.
Behind it there is usually a real complaint: work that takes too long, answers that live in one person’s head, reports assembled by hand every Friday.
The proposal on the table is a tool. A chatbot, a copilot, a platform with a pilot price. It is the one thing in the conversation that can be bought this quarter.
02 · Why the obvious answer fails
AI speeds up a process. It does not decide what the process should be.
Take a team that answers customer questions from four spreadsheets, two shared drives and a colleague who has been there eleven years. An assistant trained on those sources answers faster, and it repeats every contradiction between them with more confidence than the colleague ever did.
The second problem is ownership. A pilot is set up by whoever was keen, on whatever data was easy to reach. When the pilot ends, nobody owns the answers it gave, so nobody can say whether it worked.
The third is measurement. Most AI pilots are judged on whether people liked using them. Liking a tool and saving time with it are different results, and only one of them shows up in the budget.
03 · What we look at instead
Where the time actually goes. Not the job description — the week. Which tasks are repeated, which are waiting on someone else, and which are rework because the first answer was wrong.
Where the knowledge lives. If the right answer exists in three places with three different versions, the first project is deciding which one is true. That work is unglamorous, and it is what makes any AI useful afterwards.
What a good answer is worth. Pick one task and put a number on it: minutes per case, cases per week, the cost of a wrong answer. A tool that cannot move that number is a demonstration, not an investment.
The objection is that this delays the AI. It does, by weeks. It also means the tool, when it arrives, is pointed at a problem someone can measure.
04 · Where to start
Map the week before choosing the tools. List the five tasks your team repeats most, with a rough sense of how often each comes up.
For each task, note where the information comes from and how many sources it relies on.
The best first candidate is a task that happens often and draws on few sources. It pays back soonest, and it shows you what your data is really like.
Give the result a named owner from the start, so the tool is measured, maintained and, if it is not working, switched off.