A fun recent ship from Stripe: we trained an ML model on all of the disputes and chargebacks to give every one in your dashboard a 1 to 5 score based on likelihood that you'll win it, which we can predict given card issuer, type of dispute, etc with some accuracy.
This lets you and/or your CS teams prioritize which disputes to actually work on evidence collection for and which to either just refund-and-chalk-up-as-cost-of-doing-business or use less-resource-intensive evidence submission (e.g. with the API you could post a templated letter)
B2B software is generally understood to not be intrinsically a network effects business, but scale affords some data-driven advantages, and intriguingly it makes those advantages possible for customers who don't *themselves* have sufficient scale to do this.
e.g. You can imagine without loss of generality AppAmaGooFaceSoft see enough disputes to have teams working on Hadooping a few gigabytes of transaction data to squeeze a few bps out of their win rate, but you can't justify an engineering team on that below ~$1B of annual revenue.
"What is the ML actually doing here?"
Among other things, do you have a point of view on how likely a particular Japanese regional bank is to sustain a dispute opened in a card not present transaction against a foreign e-commerce merchant? I have a guess, but math has an answer.
Among other things, do you have a point of view on how likely a particular Japanese regional bank is to sustain a dispute opened in a card not present transaction against a foreign e-commerce merchant? I have a guess, but math has an answer.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
