Concierge booking platform

The booking flow was already short: pickup, destination, now or later, and a price. But some bookings were never accepted by a driver, and the problem was showing up everywhere at once. Customers raised support tickets about bookings that went nowhere. Drivers reported jobs they couldn't act on. Our operations team flagged it, and the booking data confirmed a pattern of declined jobs.

All four signals pointed to the same cause: the addresses themselves. Our address autocomplete, built on OpenStreetMap, only suggests a full address when the passenger starts by typing a house number. Otherwise it suggests the street alone, and lets them choose it as a pickup or drop-off point. The booking looks complete to the passenger, but a driver receives an address they can't act on, so they decline it. A drop-off point that broad also makes the fare estimate less accurate.

Make sure every booking reaches a driver with an address they can use, and gets an accurate fare, without adding a single step for passengers whose address is already complete.

Constraints

• Mobile, and in a hurry. More than 95% of bookings are made on a phone, often by someone who needs a taxi soon.
• Incomplete address data. OpenStreetMap is detailed in places and patchy in others, and we couldn't change the source.
• Five languages. Every message had to work clearly across all of them.
• Drivers decide. A driver can decline any job, and they won't accept one they can't trust.
• No added friction. The flow had to stay as short as it already was for most passengers.

Declined or failed bookings fell by roughly a quarter to a half after the step launched. Fare estimates also improved, because every journey now starts and ends at a precise address. At least three in four passengers never see the extra step, so the flow stayed as fast as it was for the majority. And with complete addresses arriving, drivers could accept jobs with confidence.

The clearest lesson was how much a small detail can matter. The problem looked like a driver issue, then a support issue, but it came down to a single missing number in the address data. Listening to all four signals together is what led us to the real cause.

Next, I'd look for a more complete address data source. The conditional step works around gaps in OpenStreetMap; better data would close many of those gaps at the source, so even fewer passengers would ever need to be asked.