If unimpeded cash flow is the lifeblood of an enterprise, a large volume of unrealized accounts payable is akin to a cholesterol buildup. Left unresolved, it is likely to put businesses in a situation where they don’t have enough capital to fund critical operations, payroll or even pay outstanding taxes. Despite investments in ensuring that payments are received on time, over 96% US businesses experience late payments from customers. According to the latest Atradius Payment Practices Barometer survey for 2017, the ratio of overdue domestic invoices in the US stands at over 45%, while 2.1% of all collectibles have to be written off completely. The financial ramifications of these numbers are clear – businesses need to reengineer the broader accounts receivable process or wither away slowly.
As it exists today, the accounts receivable process is largely labor intensive, requiring human operators to manually match invoices with purchase orders. Given how understaffed credit departments usually are, overworked employees are more likely to make errors, pushing up day sales outstanding (DSO) by as much as 3.3 days. Moreover, businesses relying on manual methods for managing collection activities spend more time prioritizing activities and gathering information than reaching out to customers about payment.
While some companies are experimenting with automated payments and electronic validation to improve collections, advances in the field of machine learning (ML) may hold the key to unlocking even greater value. The emotional intelligence that accountants and bookkeepers bring to the table is irreplaceable, but ML promises to take repetitive, redundant, and time-intensive tasks off their hands.
In essence, an ML-driven system can be ‘taught’ using supervised learning approach to read receipts, invoices information, and even collection policies. Based on predefined features like customer history, invoice details and text information, the underlying algorithm can classify receivables into different buckets created according to past due ranges (1-30 days, 31-90 days, over 90 days, disputed). As the system continues to be fed with a steady stream of data, it develops the ability to recognize overarching payment and customer behavior patterns, and ultimately prioritize accounts for collection. Bank of America is already using this technology to streamline their internal AR realization process.
Reimagine Risk Assessment
Supervised ML’s ability to prioritize invoices for collection can be further extended to flag accounts for review, specifically those that may not exhibit any warning signs on the surface. Since ML is particularly good at understanding deeper, non-linear relationships, it can leverage large datasets of invoice details, transaction history and customer details to extrapolate credit risk. To achieve this, the system assigns values to each risk driver based on historical data and uses the combined score to establish the probability of payment default. This information can be used to proactively identify and segment customers into those who cannot pay and those who don’t want to pay. In turn, clustering accounts based on probable customer behavior will help enterprises fine-tune their upstream processes.
By operationalizing recommendations generated through ML, companies would immediately benefit from boosted collections and greater productivity of their human agents. Do you agree that machine learning is the right step in the direction towards reducing DSO? Share your thoughts in the comments section below.