Technology evangelists tout AI as a general-purpose technology (e.g., internet, steam engine, electricity, etc.) that will unleash the next wave of innovations and productivity. Brynjolfsson, Rock, and Syverson 2019 argue that just like other general-purpose technologies, artificial intelligence will transform existing industries, society and spawn new industries. AI is a relatively nascent technology, and the disruptive potential of this technology remains untapped.
However, that is rapidly changing. According to the center for security and emerging technologies (CSET), in 2019 alone, the private sector invested 70 billion dollars in AI research and development . Given the level of investments and interest in AI, it might be worth exploring how disruptive companies leverage AI to unlock new markets. Understanding how to leverage AI to unlock new markets will help several SMEs and startups survive and thrive, leading to economic growth. In this blog post, we analyze how Upstart leverages AI and attempts to derive some conclusions for innovative SMEs Specifically, we derive a set of market characteristics that make them suitable for disruption using AI. SMEs can use these generic characteristics to scope for new markets to disrupt.
Founded in 2012, Upstart is disrupting the traditional lending business and, as of 28-may-2021, is valued at 12,83 billion dollars on the public market. They have developed an AI-enabled platform that connects banks and customers seeking to lend and borrow unsecured loans. Upstart claims their AI-enabled platform does a much better job at assessing the true risk or creditworthiness of the customers when compared to traditional systems such as FICO (more on this later). A more accurate creditworthiness assessment is beneficial to both the customers and the banks. The customers benefit by getting cheaper access to credit, and the banks benefit by lower default rates. In addition, banks are using the platform to digitize and automate their process of issuing unsecured loans.
Before we delve into the details, let’s first assess Upstart’s financials to understand if they are also financially successful and are scaling well. After all, we want to learn from the best of the breed. Upstart seems to have found the product-market fit and is scaling at a breakneck speed. In their recent earnings call, the management guides for a 154 – 162 percent year-on-year revenue growth. Table 1 clearly shows that the revenue growth rate is accelerating. The drop in acceleration in 2020 is a result of the pandemic.
|year on year growth rate||–||72,95%||80,65%||43,01%||162,47%|
|Compounded annual growth rate (CAGR)||85,06%||–||–||–||–|
|Operating expenses as a percentage of revenue||126,71%||121,25%||105,58%||96,96%||–|
The table also shows that the operating expenses as a ratio of the revenue are decreasing, which indicates that the unit economics is working in their favor; in other words, the company is scaling quite well. Upstart also reported a positive income from operations and positive net profit in 2020, indicating they are profitable. Now that we have established that Upstart is a pureplay AI-enabled fintech disruptor with product-market fit and is scaling well let’s explore how they have unlocked a new market in the lending business.
Credit is vital to the modern economy, and it helps improve our standard of living. For example, it enables us to buy houses, finance education, mobility (auto loan), etc. Banks are central to issuing credit, and the most critical factor influencing whether to issue credit or not is risk. In straightforward terms, risk measures the likelihood of defaulting on a loan. Since Upstart currently operates only in the US, it’s worth reviewing how the banks in the US assess that risk. Many banks in the US use FICO scores as an important input in determining a customer’s risk or creditworthiness. The higher the risk, the higher the interest rates. Most of the risk assessment model models use a combination of FICO scores and some additional variables. Upstart claims that most of these models are rules-based systems and use anything between eight to thirty variables. Also, most of these models were developed before the advent of AI. For example, FICO was invented in 1989.
In contrast, Upstart’s AI-enabled system uses several discreet models and over 1600 variables to assess the various aspects of risks such as loan stacking, income fraud, etc., to predict the odds of someone defaulting. In addition, the system continuously improves by learning from increasing amounts of data and by updates to the algorithms. The performance of Upstart’s AI-enabled system is also significantly better; they reported that their models approve 2.7 times as many borrowers at the same loss rate as the traditional banks. Furthermore, their loss rates were approximately half of those predicted by some prominent credit rating agencies for securitized loans.
Observation 1: Looking at the traditional lending business at higher abstraction levels, we observe a core process that collects data, processes the data (using models) to get the desired prediction (probability of someone defaulting or credit rating). The rule-based system then processes the prediction to deliver a digitized product, in this instance, credit in US dollars. Most of the traditional rating systems do not leverage AI and the vast amount of data available today. Companies like Upstart are changing that. They use AI and related technologies to exploit the abundance of data to deliver superior performance than legacy systems.
The traditional lending industry alienates many potential customers because of how the conventional banking industry assesses risk. Consequently, the alienated customers must either pay exorbitant interest rates or do not get access to credit. Table 2 shows that 22% of Americans either do not have a credit score or a reliable credit score, and 19% have a very low credit score. So, in total, 41% of Americans either can’t get credit or must pay exorbitant amounts to access credit. However, 80% of Americans have never defaulted on a credit product.
The consumer financial protection bureau published a report detailing the inaccuracies of existing methods to predict the risk of defaulting and how it adversely impacts the consumers. In this report, they track two groups of customers and their delinquency rates against their credit scores. Group one (PR) uses the default credit scores provide to them by credit rating agencies. In contrast, group two (No PR) challenged the data used to determine their credit ratings in court and had a favorable ruling from the court. As a result, the credit agencies did not include certain variables in determining their new credit scores, leading to better credit rating scores. The figure below shows how these two groups performed on delinquency rates for new loans.
Figure 1 Delinquency rates of customers on new loans6
From the figure above, we observe that group 2 (No PR) consistently performs better across all categories while their performance was very similar before the change in credit ratings. The difference in performance begs the question: are these legacy models efficient at predicting the true risk of defaulting? These credit-rating systems may exacerbate the socioeconomic inequalities by limiting access to credit to people who most need it.
Table 2 Break down of Americans adults and their creditworthiness scores 
|Tier||% of American Adults|
|Deep subprime (579 or lower)||13%|
|Thin or stale score file||11%|
*the figures don’t add up to one hundred percent because of rounding errors.
Moreover, the well-established financial institutions that have the resources to innovate are bogged down by factors such as regulatory compliance, legacy systems, manual processes, culture, and vested interests of the management. The inefficiencies caused by the factors above increase overhead costs, thus making it unattractive for large financial institutions to serve loans with lower margins.
The same factors that are the Achilles heel of the incumbents are the strengths of agile young firms. Upstart leverages AI and the abundance of data to build efficient and effective tools for predicting the true risk of default and automate the process of approving and delivering credit. The efficient and effective processes allow Upstart to serve the market segment that is unattractive to the large incumbents—most of the loans that Upstart provides range from 1000 USD to about 50.000 USD. The conversion rates (inquiries vs approved loans) are growing at a CAGR of 23%, and approximately 67% of Upstart’s loans were fully automated. AI to other parts of tier operations such as Marketing to drive efficiency and growth.
Unfortunately, Upstart does not report a breakdown of its customer base using the traditional customer creditworthiness scores. So, at this point, it largely remains an assumption that customers with lower or no credit scores are benefitting from Upstart’s service. Further, biases in all types of models remain a risk, especially in systems made up of discreet models that leverage vast amounts of data. The biases could have undesired consequences like discriminating against certain minority groups. More information on how Upstart works towards addressing such biases are in the no-action letter application filed with the CFPB
Observation 2: the lending industry is characterized by the presence of a large portion of customers that are underserved or disadvantaged. Additionally, the industry is plagued by high overhead costs caused by inefficiencies.
So, to conclude, industries with the following characteristics are ripe for disruption using AI:
- Data collection, processing, and making predictions are central to the core value creation processes. The above process is central to several industries. For example, think about how a doctor diagnoses diseases or how a judge decides who should walk free or go to jail.
- Leveraging vast amounts of data will drastically enhance the quality of predictions.
- Structurally ignores/disadvantages or even denies products and services to a large segment of potential customers.
- Industries plagued with inefficiencies caused by regulatory compliance, legacy systems, manual processes, culture, and vested interests of the management.
Attributing disruption of industries or unlocking new customer segments to just the characteristics above would disservice the entrepreneurial genius of teams operating companies such as Upstart. Exploring the business models of AI-enabled disruptors could be an interesting topic to explore in future posts.