Giant step forwards in collections efficiency with machine learning
The challenge
Although most Microsoft enterprise customers pay on time, over the last few years approximately 20% of customers pay late. Microsoft’s yearly revenue is over US$100bn, and the cost of capital due to late payments represented dozens of millions of dollars in 2019.
Microsoft’s existing collection process was labour-intensive, which made it difficult to scale, being based on a collection of scoring rules that were computed so the collections team could call customers, working the list from top to bottom. As the list was not dynamic, it could not reflect known customer idiosyncrasies nor changing market signals.
Creating different scoring systems for different types of customers and situations would make the process even harder to manage. Microsoft decided to deploy machine learning and to test whether a data-driven solution would improve collections performance.
The solution
Data
Microsoft used customer firmographic data, invoice data and customer contact data for the last three years, spanning millions of invoices and hundreds of thousands of customer accounts. Multiple sources of data covered invoices and payments, disputes, and customer contacts. A data cleansing/merging pipeline was constructed to create master data usable for modelling purposes. The data was thoroughly analysed to understand key variables that indicate a high probability of payment delays.
Collections
Partnering with Accenture, Microsoft leveraged SynOps — their human and machine operating platform — to use machine learning and predictive analytics to deliver significant improvements in the collection of outstanding payments. Specifically, the Intelligent Collections programme is able to predict a collections strategy from all of the historical data to estimate the likelihood that an invoice will be paid late. This powerful solution can predict with +95% accuracy if a generated invoice is likely to be unpaid, immediately triggering action and continuously reducing the days of payments outstanding. The system can be customised on client’s data for multiple industries.
Best practice and innovation
Predictions used thousands of different characteristics, including past customer ordering and payment behaviours, peer group similarity comparison, anomalous activity when comparing to peer group, known patterns associated with late payment, and changes in recent patterns. In order to perform predictions and increase interpretability, the number of variables were reduced to a few hundred, using machine learning to keep only the most important.
From an operations point of view, invoices with higher scores are prioritised for collections teams daily in much the same way Google or Bing raises the most important results to the top. Although this solution was developed specifically for Microsoft, using Microsoft’s historical data, similar solutions can be developed for customers in other industries.
By identifying potential late payers, collectors can focus efforts on them and avoid contacting customers that are likely to pay on time, saving effort and improving customer experience.
One of the main objectives was to reduce the number of contacts necessary to reach 95% of the open amount. Collectors were calling the majority of customers to meet their goals: now they can confidently meet their goals by calling a low percentage of the customers.
The model was deployed in mid-2019 and tested using a randomised controlled trial, the gold standard in statistical testing methodologies. using a stratified random selection for 20 collectors. The customers of ten collectors were assigned to the treatment group and the customers of ten other collectors were assigned to the control group.
Homogeneity was checked between both groups over characteristics known to influence performance (such as customer size, average lateness, segment, and proportion of cloud products), and verified that both groups were similar.
At the end of the test, collectors from the treatment group collected hundreds of millions more than the control group, improving collector’s efficiency by more than 10%, saving millions of dollars in capital costs per year, in addition to reduced effort and improved customer satisfaction.
Key benefits
- More than 10% increase in efficiency, saving millions in capital costs per year.
- Significantly reduced collections effort
- Improved customer satisfaction.
Revathi Subramanian
