How to confirm leads using ML
CarPrice helps people to sell their cars. All you need is to book an appointment on the website and go to the branch you selected at the time you selected. There is an analogue of WeBuyAnyCar in the UK and SellAnyCar in Dubai.
The problem
There is a number of nuances about documents and ownership details that must be clarified before a meeting so we call every person after online booking. Of course, we evaluate the conversion rate from call to deal for every call-centre specialist. There are stars who have a great CR and another whos metrics are not so good. As soon as the number of leads grows it is hard to hire such stars who could maintain great conversion.
Simple method
So we tried to search how to solve this issue by technology. First of all, we made an AB-test and started to send an SMS with the necessary information instead of calling. It dropped conversion in some segments but others started to perform better. For example, it was OK for people from big cities but not enough for rural clients. As a result, we turned on SMS even for leads from Moscow.
This decision wasn`t excellent — regularly we were getting stories from branch managers about annoyed clients who came without any documents at all. It pushed us to find a way that would be more transparent for people. The decision seemed obvious — to get an option for customers how did they like to be served (by call or by message).
As usual, we started with AB-test. Half of the people were able to select an option, the others weren`t. This experiment had shown that we would lose money in case of giving such an option.
ML instead of handcrafted rules
We discussed the problem with our analysis team and decided to make a smart-algorithm that would use ML in order to select the best way to communicate.
After a number of iterations, it was done. When CRM gets a new lead, it goes to a special API and sends them the ID of the lead. Algorithm gets an ID and crawls all the data it knows about the lead such as marketing source, car specifications, selected date and time, geo-info, gender, history on website and CRM etc. and takes all of them into account. Then it makes a decision (to call or to text) and sends it to the CRM.
Results
This project lets us drop more than half of the calls and increased gross margin per deal by 17% so we do not even save money but grow revenue.
