The Model Said No. She Said Yes. She Was Right.
Our model said no, and I agreed with it.
It was a regional distributor — auto parts, family-run, about $40 million in revenue. Thin trade file, two late payments the year before, a balance sheet that looked like it had been through something. The score came back deep in the decline band. This was at Cortera, mid-2000s, and the lender was a client running our software. I’d spent months on that scorecard. I knew exactly why it said no, and I thought it was right.
Then their credit officer — I’ll call her Linda, she’d been doing this for twenty-odd years — overrode it. Approved a limit. Put a short note in the file: “Talked to the owner. Know the business. Seasonal.”
I’ll be honest, my first reaction wasn’t curiosity. It was mild annoyance. Overrides were the thing we built the system to reduce. Every override was, in our heads, a human introducing noise into a clean decision.
Get Suneel Reddy’s stories in your inbox
Join Medium for free to get updates from this writer.
Eighteen months later that account was one of the best-paying in their book. The late payments had been two bad months after a warehouse flood, both cleared within sixty days. The ugly balance sheet was the flood too. And “seasonal” turned out to be the whole story — they ran tight in spring, flush by autumn, every year, like clockwork. None of that was invisible exactly. It just wasn’t in any field our model could read.
She’d driven out and seen the warehouse. That’s it. That was the edge.
I think about that account more than I think about most of our model’s wins. Not because the model was bad — it was right far more often than any individual officer, and the data backs that up. But because it showed me the shape of what a model can’t do. It can tell you, very precisely, how this file compares to ten thousand others. It cannot tell you that this particular file is lying about itself, in either direction. Someone has to go look.
We changed how we treated overrides after that. Stopped counting them purely as noise to drive down, and started tracking which officers’ overrides actually performed. Turns out a handful of people were consistently better than the model on exactly the files the model was worst at. That’s not a flaw in the system. That’s the system telling you where the humans belong.
I still trust models. I’ve built my career on them. I just don’t trust anyone, me included, who thinks the score is the end of the conversation.
Originally published at https://suneelreddy.com on November 9, 2026.
