Basoli Dhirai.
[ 01 / WORK ]
One policy, fair to everyone it touched
↓20%
order fail rate
↓40%
customer cancellations
The problem
Every cancellation was quietly expensive.
Cancellations were driving meaningful, recurring losses across commission, vendor compensation, and logistics: enough to make this a board-level problem, not a UX tweak.
Every cancellation has four parties, and they don't all want the same thing. Make it generous for the customer and the vendor eats it. Protect the vendor and the rider wears it. Simplify for the business and someone downstream pays.
So before writing a single rule, we mapped the whole cancellation journey; every party's version of it, side by side: what they experienced, the tools they used, and the exact points where things broke.
Getting that right meant bringing in the teams who spoke for each side. The vendor team flagged a tool that couldn't handle what we'd planned, down to the POS systems they were stuck with daily. The rider team surfaced a workaround nobody else had caught. Legal held the line on what each market would allow.
A cancellation policy has no neutral party.
The fair version made the front line's job harder, so we rebuilt their tools too.
Because we'd chosen fairness over simplicity, the refund couldn't be one flat rule. It depended on how far the order had already travelled: cancel before the vendor accepts, you'd get one amount; after acceptance and preparation, another; once it was picked up and out for delivery, another again.
Fair to everyone — but it meant a support agent now had to work out the right refund for the right stage, by hand. So we rebuilt what the agents worked with, alongside the agents themselves: new SOPs, rewritten scripts, retraining for the change, and tools redesigned to reflect the staged logic rather than fight it. In the first rollout slice those refunds were issued manually; once the logic held up, we automated them.
Listening live, as it rolled out :
A second track ran in parallel asking users why, in the moment they left.
As the policy rolled out region by region, we built a live survey into the cancellation flow itself, catching real reasons at the exact moment someone tried to cancel.
The answers reshaped things as they came in. Most people cancelling weren't trying to leave at all, they'd forgotten to add something, or needed to fix an address. Three in four reordered immediately after. That single insight eventually turned into a checkout nudge letting people edit before they cancelled.
Stuck on a problem with too many moving parts?
More projects