01 / 03

Start with a task someone dislikes

Ask a team where its time goes and you will hear specific answers. Reading supplier updates. Sorting customer messages. Writing the same explanation for a different account. Finding the change hidden in a new price list.

Those are useful starting points for AI work. You have a person who understands the job, examples of what good looks like and a way to notice whether the result helps.

A broad instruction to “use more AI” gives the team another job. A well-chosen task gives it a chance to remove one.

02 / 03

Count the whole job

Suppose a tool drafts a response in seconds. Someone still needs to read the customer’s message, check the facts, correct the draft and send it. Time all of that. Compare it with doing the job without the tool.

Include the awkward cases. A missing delivery, a refund already issued, a customer asking two different questions. If the tool produces a confident answer that takes longer to untangle than to write, record that too.

Keep a small set of examples the team can run again when the prompt, source material or model changes. A good result once is encouraging. A reliable result on the work arriving next week is what matters.

03 / 03

Keep responsibility clear

Decide what the tool may do on its own. Drafting a message and sending a refund have different consequences. The person reviewing an output needs the source information and enough time to disagree with it.

Give the team a simple way to flag a bad result. Someone should read those reports and change the workflow when the same problem repeats.

The useful measure is whether the team finishes the work with less effort and the same or better care. If it does, expand carefully. If it does not, change the task or stop the experiment. People’s time is worth more than keeping a pilot alive.