Context
At Census (acquired by Fivetran), we had a growing pipeline and a sales team that couldn't tell which accounts were worth chasing. Classic startup problem - lots of signups, no idea who's serious. I used our own product to answer the question. The approach works anywhere you have account data in a warehouse.
The problem
Every sales team thinks they know their ideal customer. They're usually half right. The traits they think matter are based on a handful of deals and a lot of gut feeling. Meanwhile, the data that could actually answer the question is sitting in the warehouse untouched.
I wanted to go from raw account data to a working fit score in a single afternoon. Not a six-week project. An afternoon.
How I built it
Let AI look at your best customers, figure out what makes them "best," then score everyone else against those traits. Instead of someone deciding "enterprise companies in fintech are ideal," the model shows you which traits actually matter.
I pulled in account data - company size, industry, what features they used, how often they showed up - and had AI compare our best accounts against our worst. The output: a score out of ten, an explanation of why they scored that way, and talking points the sales team could actually use.
Pick five to ten accounts you know well to test the scores. You'll know right away if the model is making sense. Get sales to sign off before you roll it out to everyone. If they don't trust it, they won't use it.
What I learned
- The model disagreed with our gut. Traits we assumed mattered didn't. Others we'd never thought about turned out to be strong signals. That's the whole point.
- Speed changes behavior. When ICP analysis takes an afternoon instead of a quarter, you actually do it. You iterate. You test.
- Scores need to go somewhere. A spreadsheet nobody opens doesn't help. The score needs to show up in Salesforce, on the lead record, where the sales team already works.