Instant Payments: The Application of Machine Learning
For more than a decade large firms have been able to support their suppliers with a range of supply chain finance (SCF) options. Yet, uptake of these programmes by SME suppliers has remained limited due to (a) the fact that they require an approved invoice (and approvals for the invoices of SME’s tend to take many weeks or even months), and (b) challenges in onboarding such as cost and Know Your Customer (KYC) checks. In this article, Dr Richard Carter, data scientist at Previse, explains how the latest developments in machine learning enable an entirely new approach which, for the first time, eliminates these obstacles and helps facilitate low-risk, pre-approval invoice payments.
Traditional Supply Chain Finance
Working capital is the lifeblood of businesses within the global economy. However, large buyers and small suppliers often clash over time taken to pay. Many buyers wish to extend payment terms to improve their working capital position. Suppliers, for their part, often need payment sooner to support the cash flow of their business.
Many solutions have come to the market over the last couple of decades to resolve this tension. These can generally be categorised by whether they are seller- or buyer-focused. With a supplier-focused solution, such as factoring, the supplier receives early payment in exchange for an advance of a slice of the total invoice amount. This comes at a relatively high cost, relying as it does on the supplier’s credit rating. By contrast, with a buyer-focused approach such as reverse-factoring or SCF, a supplier can receive early payment on invoices but must wait until they are approved by the buyer, and pass the stringent on-boarding processes required.
An ideal solution is one in which SMEs can receive their payments with minimal fees as quickly as possible and can be enacted by a buyer with little change in process.
Focus on Predictions, not Process
The full source-to-pay process can take several months. Once the goods have been received, a buyer will typically perform a matching process between the purchase order, order receipt, and invoice. This three-way match ensures that no discrepancies exist between the documents and that the invoice can be booked as an account payable. Invoice approval can take between a week and a couple of months extra, depending on the internal efficiency of the buyer and the prioritisation that the invoice is given. This means that within the standard process the earliest an invoice can be paid is on the day of approval.
Previse’s central innovation is to be able to pay suppliers without requiring invoice approval. Through powerful and sophisticated analytics, trained on billions of pounds of real invoice spend, we are able to predict approval and payment with such accuracy that a funder can pay the supplier instantly upon receipt.
Machine Learning and Data Science
The unique achievement of pre-approved invoice payments has come about by building a resilient and flexible technology stack which accesses a buyer’s transactional data, enriches it, then pipes it to state-of-the-art Artificial Intelligence (AI) to predict ultimate payment likelihood.
Experts in the fields of software engineering, ERP systems, data science and machine learning have combined to create the world’s only algorithmic prediction of dilution risk, i.e. the risk that an invoice is disputed and not fully settled in cash. Data-driven decision making has become a touchstone for contemporary businesses, and this is what Previse has brought to the world of B2B commerce.
The data we receive from a buyer typically comes from a highly customisable ERP system such as Oracle or SAP. Hence the way that buyers choose to use their software means that no two configurations are ever the same.
A vital step for Previse, therefore, has been to develop and build a single standard data model. Our technology normalises the clients’ own data, bringing it in line with our model without any need for the client to adjust their working practices. This allows us to ingest data from any new client and, regardless of their ERP software and configuration, materialise it into a form that is internally consistent and robust.
The integrity and consistency of the data are strengthened further by checks and balances that mitigate against any erroneously entered inputs. It is often stated that 80% of the work in dealing with data is in this manipulation and cleaning. However, we have built a highly scalable approach that enables large numbers of buyers to be onboarded without significant increases in headcount.
The core components of Previse’s system are the models that we use to determine whether we should sanction an instant invoice payment or not. The features that we use in the algorithms are increasing with sophistication all the time, but at heart, the issue is always one of binary classification: we either sanction an invoice for early payment or we do not. The outputs from the models are risk scores that measure on the probabilistic scale of zero to one, the likelihood of an invoice being “bad”. This is then compared to a buyer-specific threshold to determine our final decision to either fund or defer.
The algorithms are owned by data scientists with strong academic backgrounds. We are constantly assessing new research to determine its utility, not only in our model building but also our feature enrichment and ancillary data checks.
Anti-Money Laundering and Fraud Detection
With Previse, payment is made by a funder on behalf of a buyer. This places certain responsibilities on the funder to ensure that they detect, prevent and report any suspicious activity as it relates to money laundering and other such criminal activities.
A core component of the stack that we have built enables us to perform Anti-Money Laundering and fraud checks for our funders. Banks typically have hard-coded rules, but we have enhanced these using cutting-edge machine learning methods that are capable of detecting far more clandestine methods of using payments for illegal ends. Rather than simply checking invoice amounts against historical norms, our sophisticated model is capable of uncovering much more fine-grained anomalous values amongst all the features that we use to risk score an invoice.
This capability not only protects the program against money laundering but can also catch any unscrupulous seller that is prepared to enter a fraudulent invoice.
Across every industry, firms are realising that the data they produce and collect has enormous value. The same is true of invoice data. From analytics, through machine learning and algorithm development, data science is at the heart of Previse. We are providing new insights and driving more efficient and effective ways to do business.
Through our state-of-the-art AI and analytical technology, we are enabling firms to unlock the value in their invoice data, to get their suppliers paid instantly and build a stronger supply chain.