AI Assisted Credit Decisions

There are two prevalent opinions that surround Artificial Intelligence (AI) in current applications:

You should prepare to fight machines soon

Making of machines that start to rebel abusive humans in HBO series Westworld

It’s overrated

These types of “fails” are somewhat common within the relatively nascent chatbot industry, especially when a simple bot with limited artificial intelligence.

Today, machine learning and statistical techniques already help us automate and scale cognitively challenging tasks. But without understanding the mathematical methods that go into this, it is easy to get lost in the superficial, clickbait-style headlines. It is sometimes difficult to see beyond the hype and understand how all these technology companies and startups are actually implementing learning and optimisation algorithms, as well as how these impact the business side of things. The aim of this article is to explain the effects of AI on businesses, using the financing service Investly as an example.

The perception of Machine (Learning) and what is Artificial (Intelligence)?

We often speak about machine-made things as opposites to man-made things. This applies in many areas, such as food (manufactured vs. organic or home-grown) or genetics (bioengineering vs. breeding). However, what people often fail to understand is that the technology has been created by us, humans.

Accumulation of knowledge over generations has enabled our species to gain the skills to automate or extend our human capabilities. Humans create, operate and control these machines. Through technology, we are able to transcend our biological capabilities and enhance our humanity by using our brain power for more value-added processes beyond data entry and processing.

This understanding is reflected in all following sections.

The right kind of philosophy

At Investly, our mission is to help businesses grow by providing them with working capital — a lifeline for many companies. Even profitable businesses go bust, which is what we are here to avoid and instead we assist them in sustaining their growth. We do this by releasing cash from invoices with long payment terms (30–180 days).

Processing such invoices may sound simple, but it has proven to be more complicated in practice, because we need to:

  1. Authenticate the customer (we don’t have any brick and mortar branches for face-to-face interaction),
  2. Assess the risk of payment default (credit risk),
  3. Make sure that the goods and services have been received (contract risk),
  4. Finance the invoice (we use a marketplace model to offer the best rates on the market),
  5. Collect the payment on due date,
  6. Handle collection of defaulted payments.

When we first started, we had an ambitious goal that one day we would be able to process invoices without any human intervention, as the majority of existing providers with comparable business models largely operate manually. We wanted to fix the unit cost of processing a single invoice, providing a significant cut in financing costs for the businesses.

At the same time, we made a conscious decision that in order to do so, we would first have to get these processes running and build up expertise by processing invoices manually. This allowed us to improve the system rapidly. In other words: we had to first create the intelligence naturally by implementing best practices and involving the best expertise available.

Although the manual process has been costly in the early days, it has been essential in order to be able to teach the same thing to a machine later on. Once we have managed to replicate our natural intelligence in a computer system, we are able to start teaching the machine to improve the accuracy and efficiency beyond the capabilities of our biological brains.

This conscious decision may sound like a small thing, but it wasn’t. There are considerable time and financial costs related to software development, which this philosophy helped us come up with a process that helps us to succeed faster this day.

What is the business impact?

When you’ve reached a product-market fit, validated the acquisition channels and started to grow, you basically have two options:

  1. Hire more people to sustain the growth with existing processes
  2. Invest into development to improve efficiency per person

These aren’t mutually exclusive, but finding a right balance between these options can make a big difference.

Recruiting and hiring people requires many resources. Our average cost of hiring a person is equal to two months’ worth of their salary. Plus, it takes a minimum of three months to recruit a person that would have the right skills, cultural fit and good ability to adapt to changes. But it doesn’t stop there: you need to then train (usually 1–2 months) and manage these people. New hires usually get fully up to speed 2–3 months after starting work, which means the total process can add up to 6 months on average.

In addition, there is a marginal cost of management with every new person to the team. Communication and coordination get increasingly complicated as the organisation grows. When your customer base is growing faster than you can hire, you’ll be losing out on revenue. Don’t get me wrong: we’ve grown into a very tight-knit team over the course of our mission. We value our team and the people in it above everything else. But if we are able to teach the machine tasks that are prone to human error or are just repetitive, our team becomes happier, because they can move up the value-adding ladder.

On the other side, building technology is not without its complications either. These machines are still largely built by our engineers, who we need to recruit, hire, train and manage. However, they will be working on building tools that will have significantly lower marginal cost for managing. When built right, these tools can better withstand the demand from customers as well.

The optimal set of weights probably differ along the curve of accelerating growth. In the early days, we might have spent ¾ of effort on building the operations team and ¼ on development. Now that we have hit 40% month-to-month growth, we are spending ¾ on development and automation and ¼ on recruiting.

Who are the right people for the job?

There is a huge demand for data scientists and statisticians in most domains which see potential to improve business results via data-based decision modelling, including marketing, sales, finance, insurance and healthcare. Data analysts and statisticians typically don’t have sufficient software development skills to implement their own ideas end-to-end. This means that they rely on someone else to build the applications to implement the decisions their models make. This slows down progress.

When building the team around Investly, it has become clear to us that we need people with great product engineering and software development skills who have picked up machine learning and statistics on the side. It’s like hiring a poet with poor writing skills — you can appreciate the fresh take on things, but not much gets published.

For the reasons brought out above, it’s clear that merely understanding the data science is not enough. Many advances come from building better software architecture and reading data from different sources (ranging from .pdf scans to APIs and application databases). A complete engineer is made by combining product and software development skills with a keen interest in data analysis and machine learning. Their speed of implementation outperforms peers in most teams, which are made up of data analysts and software engineers working separately.

Investly Winter Days 2017

Having navigated through different pivots, we’ve learnt that the first few people in operations will play a crucial role in how the processes will be designed.

Before getting started

When starting from scratch, architecture design can have an influence on many future processes and business decisions. That’s why it’s important to take some time and agree on some basic principles already when starting out.

Make data available in a machine-readable form (if you can)

.pdf’s and scans are a mess to read data from, because they lack machine-readable structure. Even though it’s possible to extract data from them (which we’ve needed to do as well), it is advisable to avoid sourcing data from .pdf’s and scans in your processes, if possible.

Keep track of changes in data structure

When you start training your machine learning models, you want to have the biggest possible training set for the trained model to be accurate. However, if you’ve had to frequently change the data structure and have not kept a log of the changes, it may be that the data you’ve accumulated becomes worthless.

Need for storing vs. displaying

It is possible, but not necessary to store every single bit of information. It can be difficult to understand what one should store for the machine learning models’ future access. We struggled with this a lot in the early days, as it requires quite a bit of business understanding and learning.

A general rule of thumb: when it’s crucial to replicate a decision-making process based on the underlying data in the future, store it; otherwise just calculate it ad hoc. However, in the early days, when volumes are small and processes are rapidly developing, this has less importance.

How to get started?

The key to successful automation is to divide processes into independent modules that have some specific input and output. This is easier said than done, because in real life everything may seem to be connected to everything. But the ability to break parts of processes down into simpler modules will help to make certain problems completely machine-solvable. I’ll try to offer some guidance based on our own experiences.

Start with simple problems

Sounds stupid, but it isn’t. Typically the path toward automation starts with technically simple fixes, rather than feed forward backpropagation neural networks. Take, for example, the human brain. If you analyse a single neuron, its inner workings are quite comprehensible for us: it processes and transmits information through electrical and chemical signals via synapses to other cells, which connect to each other to form neural networks. However, these lower-level connections form more complex patterns when zooming out to different areas in the brain. Nevertheless, these individual neurons and synapses are key in enabling higher-level systems to work.

In our case, when we run background checks on our customers, we use many different sources to assess credit risk, starting from the user’s demographic data (e.g. address and date of birth), ending with their debt history with other lenders and the tax office. Receiving this data in a structured format and juxtaposed with other key information can already help our analysts save considerable time. Also, it’s near-impossible to introduce more complex learning algorithms before the basic data flow is in place. These data retrieval modules can also validate and verify data from multiple sources, which is required for making well-informed decisions.

Certain parts of processes don’t need human intervention

Switching between tasks has a high cost for the human brain (see Chapter 5 in Algorithms to Live By: The Computer Science of Human Decisions by Brian Christian and Tom Griffiths). Therefore, focus is key. If you have to switch between multiple semi-automated processes, you’ll be wasting your brain power. Every time you move from one task to another, your brain needs to familiarise itself again with the task at hand. This takes time, so the less you task-switch, the less you waste valuable brainpower. Hence, a good principle for modularisation: identify which parts of processes are just boring routines and don’t require any human intervention.

In our case, sending collection notifications is a perfect example. We have a schedule of emails, texts, snail mail and calls that we used to send out manually. We automated these notifications as well as measuring the success: did the recipient open, respond or reply? Later, the machine can carry out control group experiments to figure out what are the best times in the day to send notifications. As there is a high volume of notifications involved, having a single person carry out the experiments is impossible. However, the machine is able to figure out the optimal times quickly as well as continue to improve the schedule without intervening significantly in the future.

However, when the recipient does reply, it requires critical thinking from our analysts to evaluate the situation and find a good solution for both parties. Hence, collection can be divided into two modules: notifying and problem-solving. The former can be fully automated, whilst the latter needs specific business understanding in order to be able to find a solution.

Which process has the least amount of exceptions?

Not all problems are fit for full automation. Pitching competitions and PR publications oftentimes create an illusion that entire processes can be 100% automated. In some cases this holds true, but in many cases such an approach is unrealistic. Peter Thiel, co-founder of the fraud detection company Palantir, described their approach as follows:

The fraudsters’ adaptive evasions fooled our automatic detection algorithms, but we found that they didn’t fool our human analysts as easily. So Max and his engineers rewrote the software to take a hybrid approach: the computer would flag the most suspicious transactions on a well-designed user interface, and human operators would make the final judgment as to their legitimacy. Thanks to this hybrid system — we named it “Igor,” after the Russian fraudster who bragged that we’d never be able to stop him — we turned our first quarterly profit in the first quarter of 2002 (as opposed to a quarterly loss of $29.3 million one year before)
Peter Thiel, Zero to One

Humans still outperform machines in multiple ways. This also applies to Investly when it comes to decision-making in onboarding customers with limited data and giving them sustainable credit limits.

When designing our processes, we look at the number of different exceptions from the main use case and pick cases that can be completely automated first. Certain processes, such as assigning credit limits, can be very straightforward when the data is easily available and can be fully automated in a majority of cases. However, in some cases the data is incomplete and our analysts have to delve deeper into the customer profile in order to reach an accurate decision. The latter case would be called an exception by our machine.

Thus, we would design our credit risk assessment to give automated credit decisions when sufficient data is available (main use case module), but assessment would be turned over to analysts when expert opinion is needed due to lacking information (credit limit exception handling module). Having said that, we would try to minimise the number of exceptions over time by teaching the machine to handle more of the exceptions on its own, based on the expert opinions given by our analysts. This learning can be automated with different machine learning algorithms.

Where do we want to reach?

In his book The Singularity Is Near: When Humans Transcend Biology, Ray Kurtzweil makes a case for different levels of AI and their speed of development. The camp of “You should prepare to fight machines soon” is more concerned about General AI, which is basically machine-based human-level intelligence. Kurzweil thinks that we will realistically achieve this by 2038 (as the book was published on 2005, we’re on track with most of his predictions for 2005–2017). On the other hand, the “It’s overrated” camp speaks typically of Narrow AI, which is artificial intelligence created for a very specific purpose such as speech and image recognition, self-driving cars, fraud detection, etc.

Our vision is to create the first Narrow AI for providing working capital for businesses across the globe. As the speed of economic growth is not slowing down, we see that we have an important role to play by helping mankind in manufacturing, distributing and selling goods. With Investly, we provide business owners with the financial means that allow them to efficiently manage their growth.

As self-driving cars, we’re currently in the driver’s seat looking over the Narrow AI piloting the car. When the machine itself is not capable enough, we help it pick up some slack. All the while, we’re making progress and moving towards the back seat. Most of my generation is unaware that buildings used to have lift operators. Many manufacturers, distributors and service providers starting their businesses in 10 years’ time will be surprised to hear that just a few years ago their financial and credit decisions would have been made manually by humans.