Where rules end and review begins

Cleaning a vendor list sounds straightforward. Rename a few fields, normalize the values, and import the file. The awkward part is deciding what to do with the values that almost fit.

“USA” is usually an easy mapping. “Software & IT” is less clear. It could describe a software company, an IT services business, or a category in somebody else’s taxonomy. A plausible answer isn’t necessarily the right answer for the system receiving it.

That distinction is what interests me about the workflow around AI.

Use the rules you already have

If a system has a fixed set of allowed values, start there. Make the vocabulary explicit. Keep repeatable transformations in rules that another person can read and change.

There’s no particular benefit to having a model reinterpret the same unambiguous value every time. A lookup is easier to check and easier to explain later.

Keep the source beside the suggestion

For the ambiguous values, a model can help suggest a mapping. But I want to see what it started with, what it proposed, and which choices were available.

In Lead Data Standardizer, the list-upload tool I built and shared with colleagues, that comparison is part of the review interface. The idea is simple: a person shouldn’t have to reconstruct the original file to decide whether a proposed value makes sense.

A valid picklist value can still be the wrong choice.

Give uncertainty somewhere to go

A confidence label doesn’t do much by itself. The reviewer still needs to know what requires attention and what approving it will change.

A small, explicit decision is more useful: this input, this proposed value, these alternatives. The example in the playground makes that boundary visible. If you change the value after approving it, you need to review it again.

Sometimes a rule is enough

Audience Calendar approaches a different problem with deterministic rules. A planner can see why two sends were flagged without needing a model to interpret the schedule.

I keep coming back to the same question: what’s the least complicated way to make the next decision clearer? Sometimes that involves AI. Sometimes it’s a lookup, a well-placed comparison, or a better question.