Underwriting is at the very heart of the credit business. Gathering and analyzing information about a person or business to determine whether to offer credit and on what terms is arguably the central craft in all of finance. Given its importance, and the tendency of new entrants to underplay risk as they rush to grow market share, it is understandable that veterans in the lending business are skeptical about innovations in underwriting.

As discussed in the first article in this series, much of what characterizes a good lending business has not changed, but some things certainly have. In this piece we’ll look at what some of the most sophisticated small business lenders and “fintechs” are doing in this area, and map out how a lender looking to achieve the best possible efficiency and effectiveness in underwriting might approach this challenge.

New data sources

The first task of an underwriter – human or machine – is to gather and verify all pertinent information about a loan applicant. Increasingly, loan applications are submitted online, as we discussed in our piece on UI and UX. Any small business loan application will of course contain self-reported information about the owner and business and credit being sought. Often, supporting documentation is also required, for example tax returns and formation documents, which can now generally also be uploaded and managed online. All of this is nice and straightforwardly saves the applicant and underwriter time and frustration but isn’t particularly revolutionary.

One thing that does have the potential to change the game though is using APIs (application programming interfaces – basically services that allow different applications to communicate over the internet) to pull data from a range of sources that can be at once more reliable, more up-to-date, and easier to pipe directly into an underwriting and decisioning engine. Sometimes this will require the applicant to provide login credentials to a data source (akin to “log in with Facebook”) or to otherwise provide explicit permission, though not always.

One example that has become commonplace among “fintechs” in the small business lending world is to use a service to digitally pull bank transaction records for an applicant’s operating account/s. This is a huge improvement over manually reviewing bank statements. Other types of traditional credit information that can be gathered far more efficiently using APIs include lien searches, identity verification and other “know your customer” information, and credit bureau data on owners, businesses, or blended reports containing both. Some non-traditional types of data that can also be gathered this way include data from social networks (for example Google Places and Yelp, which can be useful for verifying that a retail business has at least some happy customers) and from payment processors, e-commerce platforms, and shippers.


Types of data that can be gathered digitally using APIs:

  • Bank records
  • Identity and KYC information
  • Credit bureau reports
  • Lien searches
  • Social network data (e.g. Google Places, Yelp)
  • Transaction and shipping records (e.g. Amazon, Shopify, PayPal, UPS)
  • Cloud-based accounting systems (e.g. Quickbooks, Xero)
  • Economic data about an applicant’s geographic area or industry

Any proprietary data a credit provider may have on customers can also be extremely useful for underwriting or pre-qualifying them. For example, Amazon and PayPal have credit programs for businesses on their platforms – on whom, of course, they have loads of useful data – and American Express and many of the credit card payment processors have similar programs.

One trick that some of the fintechs have had success with is to gather data on applicants’ online behavior as they apply for a loan. For example, some offer calculators or “sliders” on their online application form and have reportedly found some useful signal in the amount of time an applicant spends dialing up and down a loan amount using these tools. The idea is that a business owner who takes the time to understand the implications of different loan terms is a better credit risk than one who does not.

A more science-fiction flavored example of how new data can be used comes from Ping An, a large Chinese financial services firm, which is reportedly recording video of loan applicants answering questions about the loan purpose and their repayment plans and then using software to analyze facial expressions to determine whether they are telling the truth.

Innovations in credit modeling

Once all the data is gathered and validated then of course it needs to be analyzed in a way that helps make an intelligent credit decision. This involves 1) the credit model and policies; and 2) the actual process or rules engine that applies those policies toward making a credit decision. There has been a lot of genuine progress in both areas, though not without some undeserved hype.

Credit model design is of course a huge topic in and of itself, though there are also some very simple approaches that may work perfectly well for established products. For example, a small business lender may simply set minimum thresholds for time in business, sales, profitability, and debt service coverage and offer the same terms to any business that meets these criteria.

Such a basic set of rules is of little help, however, when it comes to evaluating businesses who don’t fit the standard box, and to pricing risk generally. That is a much trickier exercise, with potentially great rewards for those who can crack the problem – and failure and ignominy for those who think they’ve cracked it but have not.

Most of the “fintech” lenders that have achieved any success have done so by offering credit to segments of the market that at least appeared to be underserved by banks, and they generally did this by using new credit models, often utilizing new types of data. Some of their success has proved to be fleeting, and in some cases they certainly overestimated the effectiveness of their models. But some of their innovations are genuinely valuable, and thankfully some have made their way out into the market for use by other lenders. More on this below.

It is also probably fair to say that the buzzy concepts of machine learning and artificial intelligence (“ML” and “AI”) have been largely overhyped when it comes to underwriting, at least for small businesses. It is true that ML techniques have achieved intriguing breakthroughs in some areas, such as game playing, language translation, and image recognition. A key thing that these areas have in common is a large volume of accessible data with consistent patterns that can be used to train the ML algorithms on. This simply does not exist in the world of small business loan underwriting, at least not without major qualifications.

As for AI, this is a term that is used and abused to refer to a lot of different things. A broad definition would have to cover the thermostat that activates the compressor when your refrigerator gets too warm, and would certainly include autopilot systems used in every commercial airplane. A useful if slightly tongue in cheek definition is “anything a computer does that used to require a human.” Sure, AI in the lending world does hold the promise of automating most or all of an underwriting process, and is already doing just that for many consumer products. But the reality is there’s a lot of engineering under the hood and for the most part these AI systems are mostly old-fashioned software that requires the underlying logic at each decision node to be programmed in, although the tools for doing this and the dynamic ways the systems can be deployed are indeed getting ever more sophisticated.

One caveat to the above is that AI and ML do hold substantial near-term promise for facilitating some of the lower-level functions that support underwriting. For example, ML and AI-enhanced image recognition software is getting very good at digitally analyzing documents and images that previously required manual review. This is also extending into areas like China’s Ping An using an AI-driven borrower interview process, as noted above. Though for now the uniformity, scale and economics needed to make this of real value are largely absent in the business lending world.

The promise of AI-powered decision engines

What autopilot is for an airplane, a decision engine can be for a lender. Sure, call it AI if you like. What has changed here since underwriting software was first being programmed in the 1990s, if not before? A simple but hugely labor saving – and error avoiding – innovation is simply in the way that decision engines are now able to connect directly with application handling systems on the one end and loan boarding and servicing systems on the other.

Traditionally, all business loans, including to small businesses, would involve the manual preparation of a credit memo containing all pertinent information about a prospective borrower. This memo would get distributed to a credit committee who would meet to discuss it before making a credit decision, which usually required consensus. Naturally, this requires a fair amount of manual work by highly trained (and generally well compensated) professionals. This means it’s an expensive process and only makes economic sense for large loans where the gross margins make it affordable. This is a big reason why many banks have been reluctant to lend to small businesses.

With a more automated decision engine, where most or all of the necessary inputs are piped in digitally, the data gathering and processing burden is hugely reduced, which can dramatically shift the economics of smaller loans for the lender. In the consumer market, many credit applications – notably for credit cards – have been almost entirely automated for years. Some fintechs are doing this for certain small business loan products as well, at least for clear approvals or declines. In most cases, however, small business lenders still want a degree of manual oversight for approvals. To support this, even the most cutting edge decision engines serving this market generally allow easy configurability to let lenders choose which steps to automate and when to require manual review, which can also be subject to specified filters or thresholds. This is LendingFront’s model.


Modern lending systems make it easy to modulate and automate underwriting requirements based on the size and complexity of a potential loan


Modernizing your underwriting operations

How a small business lender should approach new underwriting technology will of course depend on which products and markets the technology will be used for, what current systems and policies will need to be worked with (and which can be replaced or worked around), and what proprietary data is available. A good place to start is simply to review your current and planned underwriting workflow, making sure to capture all inputs and data sources available for consideration, the logic underlying each decision node, and the party responsible at each step. This makes it easier to visualize which components are essential and which could potentially be removed, streamlined or automated, and what that would entail. It will also equip you to be a savvier customer when you start to explore which tools and services you will need, internal and external.

Once that is done, the next tasks are to define your credit policies (and any underlying credit models or scorecards) and to select a platform for implementing these policies (i.e. a decision engine). These two parts should probably be handled in parallel since each has important implications for the other. There are at least three major approaches, which we’ve outlined below. (Full disclosure: LendingFront supports the “hybrid” and “proprietary” models but not the “fully outsourced” one.)

Approaches to credit model and decision architecture

  • Fully outsourced. There are now a number of “lending as a service” companies in the market that will fully underwrite small business loans, usually for a share of the economics.
    • Decision engine: Generally these models assume full control over the entire process from application through origination, including the decision engine.
    • Pros: Potentially easier and cheaper (at least initially) to get up and running. Limited risk, depending on how economics are shared with service provider.
    • Cons: No source of differentiation or ability to customize offerings. May involve other restrictions on the customer relationship. Limited economic upside. Limited visibility and control over credit model and decision logic, which may restrict ability to make adjustments as the market or internal objectives change. May also raise questions with regulators who expect lenders to have ownership of credit and risk models.
  • Hybrid outsourced. There are also service providers that have developed models that use a combination of their own credit bureau-style data with information on applicants provided by their lender clients.
    • Decision engine: Although the providers of these models may perform the core credit analysis, they still rely on other systems to manage the actual decision process and the surrounding application and origination workflow. The more technologically sophisticated providers of these models make integration with these other systems easier by offering API access.
    • Pros: Well established, robust scoring models are available. Fairly quick and easy to implement. Can be cheaper in the short run than building your own, with better long-run economics than a fully outsourced model. Modest flexibility and control, though these setups can potentially be augmented by or swapped out for a custom model with minimal disruption.
    • Cons: Modest differentiation and limited ability to customize the model based on proprietary customers or data. Modest ability to make adjustments as the market or internal objectives change.
  • Proprietary model. The only way to ensure full visibility and control over credit decisions is to have your own model. If you lack the necessary expertise in-house, this can be done by hiring a specialized consulting service. There are a number of firms in the market that have deep experience and well-established procedures for this. (A loan software provider like LendingFront can also refer and help evaluate qualified firms.) These firms can work with you to determine which data and records will be needed to develop and backtest a model, and to source this data if it is not available internally. Generally these models will incorporate and weigh a mixture of internally gathered and externally sourced data to assign a risk score to a prospective borrower, which can then be used to make credit decisions, determine when (and what) additional information may be required, and to set loan limits, rates, and terms. Often going through this process will reveal some types of data to be dispensable (which can help streamline the entire process) and turn up new data sources that may provide valuable credit signal without undue cost or burden. Banks in particular are notorious for frequently having a treasure trove of underutilized proprietary data on small business customers.
    • Decision engine: To put a proprietary model into use will of course require a set of procedures or decision engine. Traditionally this was done manually – and in many cases it still is. Now however there are well established software platforms that can automate as much of the process as you choose while still enabling full visibility and control. This is LendingFront’s model.
    • Pros: Full visibility into and control over the credit scoring model, supporting and resulting data, and decision logic. Potential for strong proprietary advantage, which can grow over time as more data is gathered and the model is refined. Ability to make adjustments at all levels (model, product, decision logic) as market conditions and internal objectives evolve. Although requires up-front investment, promises the best long-term economics. Can easily be augmented or swapped in and out for a hybrid model (see above).
    • Cons: More up-front cost. Danger of misjudging risk if process is not well managed.

To summarize the major trade-offs to be considered here:

  • Up-front cost v. long-term economics
  • Control and proprietary advantage v. speed to market
  • Borrower-level flexibility v. speed, consistency and auditability (this can be particularly important if you work with external funding sources, syndication partners, or loan buyers)
  • How to make optimal use of the skills and attention resources of your lending staff (e.g. free them from redundant or low value-added tasks so they can focus on engaging with customers and dealing with exceptional situations where their judgement is needed)

As is so often the case, there are no simple right answers. That said, there are almost certainly better solutions for your organization, and you will be more likely to arrive at one if you and your team systematically think through the possibilities with a healthy combination of skepticism and open-mindedness.

We will explore further the topic of underwriting and how it fits in with lending operations generally in later installments in this series. For example, upcoming articles will focus on product mix, loan origination, and overall process automation. Stay tuned and thank you for reading!