Manual vs. AI driven underwriting: Finding the middle ground with J.A.R.V.I.S

By Tushar Garimalla Globally, credit underwriting has evolved to a state where most commercial loans are underwritten fairly manually while consumer loans are underwritten using some form of scorecards or automation. Small business finance is where the challenge and the debate lies, since loan sizes are small enough to necessitate low operating expenses and automated underwriting processes, but complex enough to pose significant credit losses if expert underwriters are not involved in the decision. As a lending company starts thinking about their underwriting philosophy and building out credit processes, there is a clear design choice to be made. Do you hire lots of expert underwriters who can sniff out potentially bad loans or do we build state of the art artificial intelligence to completely automate the underwriting process? I think of the choice here as trying to hire and manage a large number of Sherlock Holmes or building a HAL9000! It might be great to hire many Sherlock Holmes, but my guess is that they either don’t exist/are hard to find, are probably really expensive, get bored quite easily and are unlikely to be able to work with other Sherlocks. On the other hand, building a HAL9000 is extremely challenging, teaching it ethics is probably harder and training artificial intelligence has a large balance sheet impact in the learning phase. Given these choices, the obvious question is – is there something else that works? While there are small business finance companies in India, and globally, that are attempting the Sherlock and the HAL9000 approach, there is an interesting hybrid approach that one could call J.A.R.V.I.S. (which stands for Just A Rather Very Intelligent System) This approach entails building automation that eliminates low risk, low value work and focuses expert judgment on very specific and precise aspects of loan decisioning. Although, eliminating low risk and low value work is obvious, appreciating the importance and cost impact of focused human judgement is critical. As an example, lack of tax compliance by a business is typically correlated with high risk on loans. However, due to clunky regulations in a certain city of India, small businesses are willing to pay the tax penalty (lower than interest on a business loan!), and use the payable tax as leverage. While that behavior is pushing the boundary with the tax department, it is a sign of a fairly astute business sense which actually brings down the risk in the business. It would cost a lot of money for an algorithm to “learn” this behavior since the lending company would have to make loans to tax defaulters across cities, wait for defaults to happen everywhere, except the city in question, after which the machine would be smart enough to work independently. More likely, nobody would seed such a test and the machine would miss out on making good loans in a city with this astute business sense. The presence of an expert underwriter though, makes this fairly easy since they are able to approve loans in this city that machines would typically decline. In general, experts allow for test designs that are more focused, cater to local business nuances and are thus a lot more targeted and cheaper. There are many such unique insights that underwriters provide and machines don’t have to learn, which makes the choice of building J.A.R.V.I.S. an intuitively better choice, however a key and much ignored aspect of this process is culture. It is crucial to understand the people that make up the team that is tasked with building J.A.R.V.I.S. Since everybody hired to build J.A.R.V.I.S. is an expert in their field and nobody has built J.A.R.V.I.S., it is imperative to bring their knowledge to the job, but not their biases. The team typically comprises of a hacker or techie, who has no clue what financial services are and thinks of computers as easier to manage than people; a decision scientist, who has built automated credit algorithms for a consumer finance business and has only heard of human underwriters as people who resist any and all algorithms; and an expert underwriter, who has managed large portfolios without automation and is constantly torn between “am I working towards making myself obsolete?” and “these kids have no idea what they are doing – why did I sign up for this?” Given these varied backgrounds, it is crucial to focus the energies and expertise of each team member towards the same goal. Defining a clear vision of the end state for the company, with an end state job description of each team member, is important for people to know that they will have a role in the company they build and it is a role that they would like to be in. Getting the team to understand that changing their view of the “other” and working with them, rather than in spite of them, is the only way to succeed. Culture is very abstract and is typically spoken of in esoteric terms. In my mind the culture that works is one that is implemented tactically. Keeping relevant team members in the loop on all decisions to ensure well rounded choices are made; spending time in meetings to explain concepts without condescension (because everybody needs time to learn); eliminating silent disagreement by seeking questions and thoughts from each team member (not just the vocal ones); disagreeing and debating publicly and openly, rather than behind closed doors (to make it acceptable to disagree); and most importantly, letting go of people who do not fit in this journey, is vital to the making of J.A.R.V.I.S. The author is VP Decision Sciences, Capital Float