I Asked AIBack

Ten checks before you spend a euro on AI.

Written for an owner or operations lead of a company with 10 to 200 people. No tool names; the checks outlast the tools.

1. Name the one job. One task, done weekly, by a named person, that takes more than two hours. Not "customer service"; "answering the twenty most common order questions".
2. Count the hours it costs today. Multiply by the loaded hourly rate. That number is the ceiling on what the AI version may cost in a year.
3. Collect twenty real examples of the task done well, with the inputs and the finished output. If you cannot find twenty, the task is not routine enough yet.
4. Decide who checks the output and how long the check takes. If checking takes as long as doing, stop here.
5. Write the failure you cannot accept in one sentence (a wrong price quoted, a customer's data leaked, a legal promise made). The system must be built so that failure is impossible, not unlikely.
6. Run the twenty examples through a general tool first (a chat assistant, by hand) before paying for anything bespoke. Score each one pass or fail against the human version.
7. Pass rate under 80 percent? Fix the inputs and the instructions, not the tool. Most failures are missing context the human had in their head.
8. Put a monthly spend cap on every AI account on day one. Agents that retry in a loop are the usual way a small bill becomes a large one.
9. Keep the human in the loop for the first month, approving every output, and measure the time saved. Remove the approval only for the output types that passed every time.
10. Review after 30 days with the numbers: hours saved, cost, failures caught. Keep it, widen it, or drop it. Do not start a second job until the first one has survived a month.

If you want a second opinion on your one job, reply to any I Asked AI email and I will answer.

Botzone Limited, Ireland. Privacy. Some links on this site may become affiliate links; if they do, the tool costs you nothing extra and I say so next to the link.