TRAINING AI WHEN EXPERTS DISAGREE

Your gran thickens her gravy with flour. Your aunt swears by cornflour.

They’ve both made perfect gravy for forty years. Eventually, at Sunday lunch, you ask which way is right.

And then you’re reminded that your uncle can’t eat wheat. Ah.

Work has its own gravy moments. Two experienced reviewers rate the same report differently.

Two planners put the same changeover on different shifts.

When you’re training an agent to do that work, the reflex is to settle it: pick one answer and call it correct.

Sometimes that’s exactly right. Some disagreement is just noise, and worth cleaning up.

But some of it is knowledge. One planner knows the fitter is on days this week. One reviewer remembers the supplier changed. And some is genuine uncertainty, which should stay open until there’s evidence to close it.

You can’t tell which until you ask why.

An apprentice who learns from both experts works out, over time, when each approach fits. Mark an agent against a single answer, without the reason behind it, and you may teach it that a valid alternative is wrong.

So before anyone writes the answer key, find out who’s coming to lunch.

Happy Monday.