The Column That Tells You When the AI Is Wrong
Situation
Every night, a small automation reads the new comments from customers who gave us a low score on the survey. It sorts each one against a codebook, suggests where it should go and who should own it, and drafts a first reply.
A person reads all of it. These are the customers telling us what we need to fix, and following up only works if a human actually listens. I wanted AI to speed up the work so whoever picks up a comment starts with a draft instead of a blank email.
But I kept getting stuck on a different question while I was designing it. How would I ever know whether the AI was any good at this?
The AI Move
I made the sheet answer it.
Next to every column the AI fills in (the category, the routing path, the owner), I added a column for the human's final answer. The AI's suggestion stays where it is. What the person actually decides goes in its own place.
I built that in from the start, and it did two jobs. It let me compare the AI against what a person would really do. It also gave me a way to see whether people were filling in their side at all. If they weren't, I'd know.
Then I tested it on real rows before I trusted it. Twenty comments, two batches. After each batch I read what came back, gave feedback, and the skill got rewritten. Four times in total.
The emails were the problem. They sounded like AI. All the wording you'd expect, nothing a person would write to someone who was upset with us. Part of the cause was that the drafts kept handing the classification back to the customer, as if the email were reporting on its own sorting. So I pulled the email draft apart from the codebook. The codebook decides where a comment goes. The draft no longer works from the categories.
I also kept it unscheduled until I'd run it by hand once. With detractors, I wanted to be reading at every step.
The Shift
When the first drafts came back, my reaction was "oh shit, there's no way we can send these automatically."
A customer would read one of those emails and know within a sentence that it was AI. For someone who just told us we let them down in some way, that is the last thing we want to hand them.
I'd started out thinking about speed. That moment changed what I was building. The goal became a process that gets better every week, and a team that trusts what the AI hands them a little more each time, for the right reasons.
We still review every email that goes out, and we still rewrite most of them to sound more like us. The rewrites are getting lighter. I couldn't tell you that without the second column.
The Pattern
Without the human column, "does the AI seem good?" is an impression. I'd have formed one in the first week and stuck with it.
With it, every row is a small comparison. Here is what the AI suggested. Here is what a person decided. Most of the time they match, and sometimes they don't. The rows where they differ are the useful ones, because each one points at something specific: a category the codebook handles badly, an owner the AI keeps guessing wrong, a draft that still reads like a form letter.
It also does something quieter. The column is a place a person has to put an answer, so it keeps a human in the loop whether or not anyone is feeling diligent that day.
I added that column during the first design pass. Adding it later, after the AI had been running for a month, would have meant a month I couldn't measure.
The Implication
If you're building anything where AI suggests and a person decides, give the person's answer its own column from day one. Two columns: "AI suggested" and "Human decided." Never let one overwrite the other.
Then look at where they differ.
What I'm Testing Next
The human column is what I'm watching now. Where the answer differs from the suggestion, and how the final emails differ from the drafts, is what feeds the next version of the skill. I'm curious how much lighter the rewrites get, and whether the emails ever stop sounding like AI to the people who read them.
Comments