Good morning. The credit module went live last week with three pilot distributors. It took 18 months from the first spec to the first dollar facilitated. Of those 18 months, roughly 10 were spent on a problem I didn't expect.
The unexpected part
I assumed the bottleneck would be the underwriting model. It wasn't. The underwriting model was ready in month four; we'd trained it against two years of order data and the calibration was clean.
The bottleneck was customer trust in the model. Distributors who had been extending credit on intuition for fifteen years did not want to hand the call to a screen, no matter what the screen said. So we spent ten months building the parts of the product that have nothing to do with credit: a customer-success motion that walks the distributor through their first three approved and three declined deals; an override workflow that lets the distributor say no to the model when they want to; a post-mortem dashboard that reviews every override and shows them where their intuition won or lost.
By the time we shipped, the model had been the easy part for a year. The hard part was teaching a distributor to trust a tool they didn't build.
The takeaway
I think the lesson here applies broadly: the underwriting / matching / ranking / recommendation model is almost never the rate-limiting step in B2B emerging-markets software. The model is necessary; it is not sufficient. What ships is the workflow around the model. Build for the override.
This month elsewhere
- New essay: The Ukrainian export thesis — why Made-in-Ukraine isn't a brand campaign.
- Ep. 12 of the podcast — shipping from Lviv to Lagos with Olena Hrytsenko.
- Two UPL assignments mid-month. Knee held.
Reading
Three books I'm re-reading right now: The Goal (Goldratt), Stoner (Williams), and The Gates of Europe (Plokhy). The first two have nothing in common except that they both improve every time I open them. The third is required reading.
Reply with what you're shipping this month. I read everything.
— Oleksandr