American Express (Amex) is a globally integrated payments company, providing customers with access to products, insights and experiences that enrich lives and build business success. And inside the company, the Amex Credit Fraud Risk business unit's mission is all about minimising credit and fraud losses while promoting business growth and delivering superior customer service.
Nothing about this will surprise you so far, we're presuming. What may: while the financial services industry uses digital for just about every process imaginable, there's one surprising remaining exception-the commercial card underwriting process, which to you and me is ‘Are you going to lend my small business any money?'
In a lot of Europe, this process is still completely manual and takes an underwriter a good chunk of time to complete. That's because, unlike when you and I ask for a new credit card, a lot of due diligence needs to be performed on a commercial entity's application, and a lot of the data sources and inputs for that are often in multiple locations and formats, on paper, or just need proper scrutiny to evaluate fairly.
For the industry as a whole, Amex representatives inform us, providing access to credit has often historically taken anywhere from two days to 30, and is very manually intensive. The business has to provide its financial data or bank statements and then an underwriter needs to sit there, take in a lot of complex documentation, review each document one by one, compare complex/disparate data in many different formats, and then provide the applicant business with a decision.
That's probably all fine in normal times, but in case you didn't notice, we haven't exactly been living in those for the past year or so. With so many small businesses still reeling from lockdowns and often desperate for lines of credit to help, American Express says it has stepped up to help by using extensive machine learning to speed up commercial card writing and make much faster approvals (or not).
Accurate risk-analysis and real-time data
What we are talking about here in terms of ‘credit,' by the way, is not a loan but the issue of business American Express charge and credit cards that a business can immediately start to use to acquire goods and services. The company stresses that approving posting of such cards is only ever based on accurate risk analysis and real-time data, supported by extensive automation of key underwriting processes, removing a lot of delay. Or as Amex's London-based VP and Chief Credit Officer, Global Commercial Payments, EMEA, Jill Zucker Sheckman, told us:
The business problem here is the ability to extend credit to commercial customers-everything from SME up through to large and global-and how we have been able to leverage AI and machine learning to reduce the onboarding time and the time that it takes to make decisions. And for businesses, especially in the light of COVID-19, there's a need for much quicker access to credit.
This is confirmed by her colleague in the American Express Credit and Fraud Risk unit, Radhakrishnan G, VP of Commercial Payments Decision Science and Machine Learning at American Express globally, who says::
To properly and fairly understand the risk of a business, you have to ask them to give their transaction records and bank statements and what other credit bureaus have worked with, and so on. They understand this can be best done if we get the right kind of data to help us make the right decision, but there's a lot of complexity here. We also want to make it a process that is a great experience for the customer.
Sheckman is responsible for the credit risk management of all commercial products for international (i.e., non-US) customers, while Radhakrishnan is responsible for all commercial and merchant risk modelling and data science strategies on a global basis; both their teams work together to extend credit to commercial customers outside the US "in a responsible manner".
But has trying to speed up the process through technology worked? Amex is very sure it has. Sheckman says:
Machine learning absolutely has helped us extend credit, especially right now when businesses are still trying to recover and need fast decisions to do that. So then instead of waiting 30 days to find out how much credit you have as a business owner, you now get a more or less instant, and accurate, decision.
Overall, use of machine learning at Amex has also led to a 20% to 30% improvement in our risk models, plus an overall reduction in cycle time - so better and quicker decisions. We think that is extremely helpful for both the SMEs and the large businesses we serve as we all go forward to recovery.
What's also clear is that this business success has not been achieved through a wholesale cull of underwriters, all now replaced by robots: Sheckman confirms that and says:
Yes, once we had people sitting in our credit operations team taking in data or documents, but now the customer can self-service in some cases, or a salesperson might be able to assist them and tell them they can call our credit operations team, ask for a line increase, and can get one potentially in a matter of seconds.
But there are still many underwriters here, and there is still lots of opportunity for such specialists in American Express: there are many other things that those people can do, especially servicing customers. This is just a shift in resources, and in most cases, underwriters have just been just redeployed.
Amex has committed to AI to automate a good part of what had been one of the few remaining manually intensive processes in its business. Sheckman and Radhakrishnan confirm that conventional technologies such as OCR (optical character recognition) and document management sit at the start of the automation process, rendering all the necessary paper (or electronic) information credit risk decisioning needs into the form the AI software needs to ingest it in, as Radhakrishnan confirms:
We've spent close to a decade to ensure what the right information is we need to gather from the business owner and how to capture it efficiently so then we can derive what we need from the data to be enabled to let the AI and ML come in to make the decision most beneficial for the customer.
But the team are not really willing to elaborate on, it turns out, what happens in the sausage-making machine once all that goes in. Radhakrishnan was unwilling to say what tools (open source? R? Python? Machine learning models or frameworks?) he and his team apply to the formatted data. He says:
My team has the specialized technical knowledge, they understand the business, they understand our products, they understand risk. We apply data science, which is a mix of business intuition and AI techniques to make a decision. But what kind of data we ask for and what software we use on it is a matter of competitive advantage. We do use industry standards, but we have heavily customized it for our specific use cases.
It's an answer that is a little frustrating in terms of detail. But surely, we can agree that if a brand sees huge value in its technology, we probably have to be realistic in terms of asking if we can root around too much in the treasure chest to see what's in there?