Six weeks to rebuild trust across 20% of at-risk ARR.

Apartment List didn’t have a billing product. Invoices were scattered through email threads, support tickets, and whichever spreadsheet Finance had made that month. Partners only learned what they owed when an invoice arrived. Roughly $5m in ARR had already been flagged as at risk.

The company was measuring the problem in support hours, because that is the half of it with a queue attached. Research said the expensive half was somewhere else: partners who couldn’t predict a bill were spending less on purpose, and a partner who quietly reduces their budget never files a ticket. I argued the target should be suppressed spend rather than deflection, and that changed what was worth building.

With six weeks, one PM and one engineer, the obvious answer — full billing visibility — wasn’t buildable. I prototyped it anyway to find that out, and the prototype did two things: it set the boundary of the MVP, and it gave leadership the evidence to buy that scope instead of building it. Inside the boundary I went after three specific barriers.

Billing questions moved from 45% of partners a month to 14%. Self-serve upgrades moved from 1% to 8%. A fifth of the at-risk ARR, roughly $1m, was recorded as recovered.

Charges arrived without warning, with nothing to check them against. Partners couldn’t predict a bill, see what drove it, or change a plan without calling an account executive.

That cost showed up three ways: support hours, slow collections, and suppressed spend. The last one was both the least visible and the most expensive.

“It made me want to spend less” is the sentence that moved the target. Everyone was counting the support cost, because support is the one surface with a queue attached and a number beside it. A partner who reduces their budget because they can’t predict a bill files no ticket, calls nobody, and appears in none of the figures the company was managing to.

That changed what success meant. Deflecting tickets was the easier outcome to optimise for and the cheaper one to prove; getting partners confident enough to spend was the larger one, and the two want different products. Working back from there, the confusion was doing its damage in three specific places.

Naming them separately is what let me argue for a scope that wasn’t a billing system. Three barriers can be attacked one at a time by one engineer. A billing system cannot, and every plan that started from the system rather than the barriers ran past six weeks before it reached anything a partner would notice.

Three mechanisms, one per barrier. A ZIP-code rates calculator with real-time totals, so a cost is knowable before it is committed to. Itemized invoices showing what drove each charge, so a bill can be reconciled against something. Self-service upgrade and premium-placement paths inside the existing workflow, so changing a plan doesn’t require a phone call.

I cut full billing visibility, auto-billing alerts, and the differentiating features from the original scope. The cut is the more informative half. None of it went because it was unimportant — full visibility was important enough that the company bought it.

Three things existed afterwards that hadn’t before. Finance could forecast without manually reconciling upgrades, because the upgrade and the invoice had become the same record. Upgrades arriving without an account executive went from a rounding error to 8%, which makes self-serve a revenue path rather than a support cost. And the company owned the scope this team couldn’t build, bought on the evidence of a prototype that was never going to ship.

The ticket reduction is the number that gets quoted and it is the least interesting of the three. Deflection saves hours. The other two changed what the business could do next.

Support was the only instrumented surface here, so support tickets were what the problem looked like from the inside. Every number in the room came from a system that can only count people who complained, and the partner quietly cutting their budget is invisible to all of it.

What I do differently since: before accepting the metric a problem arrives attached to, I ask which behaviour would leave no record at all. On this project that question was worth more than the research that followed it, because it moved the target from deflection to spend before any design existed — and two of the three mechanisms that shipped are aimed at a cost nobody was reporting. It doesn’t always change the answer. Here it changed which of two products was worth the six weeks.