Would your company invest in AI technology if doing so would increase profits by 34%?
Earlier this year, Dr. Christopher Amaral’s research group found that dealerships that applied machine learning to their loan decisioning software increased profits by over one-third. The downside? Riskier borrowers were charged more due to a higher interest rate.
Though machine learning lending software has proven profitable, balancing accuracy with transparency – and then marrying transparency with efficiency – can be challenging for machine learning & AI developers. Explainability is also an ongoing issue for AI developers, especially regarding sensitive financial data.
Refining machine learning for lending purposes has meant understanding what outcomes lenders prioritise – speed and accuracy. But, in tangible terms, machine learning’s role in lending means imbuing lending technology with higher accuracy, faster speeds and helpful capabilities like sophisticated fraud detection.
An important assumption to dispel is that AI produces value on its own.
As this viral meme (originally from eatliver.com) illustrates, the presence of AI has become somewhat of a gimmick. However, in lending circles, AI needs to have provable worth.
AI and machine learning algorithms have not had a significantly positive impact on the lending industry until the release of ChatGPT 3.5 in November 2022. As Director of Product Sam Goodacre notes in an interview with Credit Strategy, ‘Until last year when ChatGPT 3.5 was released, everything in AI was fairly underwhelming up to that point. Voice-activated AI like Alexa initially had some novelty but no real practical applications in lending’.
However, the rise of large language models (LLMs) has introduced several new applications. The ability of LLMs like ChatGPT to ‘read’ information contained in documents and images and then generate an appropriate response introduces a plethora of applications. Examples of these applications include:
The ability of AI to understand the information in financial documents (like bank statements) has been hard-won. The variable structure of bank statements – including page formatting, language, logos and so on, confounds lower-grade data extraction technology. Rule-based technologies struggle with exception handling, requiring human intervention – the exact opposite of what automation is for.
The addition of AI algorithms has augmented traditional data extraction technology, making it capable of 'reading' documents like an analyst. Analysts understand how the information in financial documents interacts when they parse them. For example, an analyst would understand that an applicant’s credit score strongly indicates their default risk. Now, AI does, too.
The result is more accurate and sophisticated data extraction. AI algorithms can scrape the essential information, perform basic calculations and then present the output for the analyst’s review. Analysts can repurpose those hours for more high-value activities (such as learning and development, nuanced risk assessment, relationship-building and more). McKinsey aptly describes this productivity realignment as ‘simple and agile’.
A sophisticated loan underwriting system will leverage the extracted data to trigger certain functions. For example, if the extracted data indicates that an applicant is eligible, it will automate the creation and distribution of underwriting documents.
AI-powered chatbots offer greater personalisation and a higher response volume than their traditional, rule-based counterparts. For customers looking for assistance with a loan application process, AI can alleviate the burden of manually responding to each query. However, when necessary, AI-powered chatbots can defer conversations to human agents.
Like previous chatbot incarnations, rule-based fraud detection systems are rigid, recognising patterns of fraudulent activities yet requiring time-consuming calibration to learn new patterns. In contrast, AI-powered fraud detection systems can analyse vast amounts of data to identify subtle anomalies and potential threats – offering more sophisticated and timely fraud prevention.
Of course, this list of AI applications is not exhaustive. AI for lending is taking increasingly innovative approaches. Let’s explore what the may future hold.
As lenders increasingly adopt AI-powered automation, the future of AI in lending can take many forms. Let’s discuss a few potential capabilities for AI-powered lending.
When used specifically for loan decisioning, AI’s biases must be managed to deliver a fairer service.
For example, there has been a gratifying volume of in-depth research dedicated to how algorithmic bias in machine learning algorithms can discriminate against specific groups. Yet, it’s not enough for the research to be conducted – lenders must implement measures to prevent bias. For example, in 2021, research by Markup revealed that intelligent algorithms demonstrated a systemic inequity, resulting in significantly higher loan denial rates (i.e. 40-80%) for minority applicants.
Other than establishing guardrails and human oversight, one way to maintain fairness in lending is to leverage alternative data sources.
AI-powered systems can incorporate alternative data sources for fairer and more accurate recommendations and decisions.
Examples of alternative data sources might include:
There has been speculation that AI-powered algorithms might use social media as an alternative data source. However, alternative data sources must be manually reviewed to ensure that only non-invasive, contextually-relevant alternative data sources make their way into the loan decisioning process.
It is not unreasonable to expect the functionality of AI-powered loan decisioning systems to improve, as LLMs’ capabilities are constantly enhancing. Learn more about the capabilities and limitations of current LLMs here.
McKinsey estimates that, by 2030, all underwriting for insurance will be reduced to seconds, with the entire process transforming from what we know. Likewise, loan underwriting will become compressed into an accurate, seconds-long iteration.
The loan decisioning system can ‘read’ more documents simultaneously without compromising performance. Parallel processing means that loan decisioning systems can process multiple documents simultaneously across multiple processors or servers. Often, these systems are cloud-based, meaning they’re scalable.
Errors are cleansed from the dataset, with AI algorithms quickly catching sparse errors and then correcting them. Where necessary, third-party sources can enrich the data (e.g. credit bureaus or government databases), which further validates its accuracy.
Of course, waiting for the capabilities of AI-powered lending systems to improve should not prohibit interested parties from exploring how AI can improve their lending processes.
DF Capital has provided over £2 billion of funding since its inception in 2016. One of the side effects of the organisation’s rapid growth is needing to process thousands of invoices per year – manually.
DF Capital’s priority, however, was maintaining in-house autonomy over document processes. They turned to intelligent document processing (IDP) to extract the required information from their invoices. After a Proof of Concept (PoC) that demonstrated 100% accuracy, they turned to Evolution AI’s invoice processing solution. Since its implementation, Evolution AI has saved 95% of processing time for DF Capital. Rachel Taylor – Head of Continuous Improvement at DF Capital – commended Evolution AI’s customer service, commenting that our team was "...really supportive – hand-holding us every step of the way."
- Read the full case study here.
Evolution AI has worked with various lenders – including Novuna Business Finance and YouLend – to cut costs and streamline the loan underwriting process.
Our automated data extraction and analytics platforms allow faster and more informed decision-making. We guarantee complete accuracy when using our managed service (including human annotators).
Get in touch with our financial data project team to learn more about what AI-powered lending can do for your business. Book a demo or reach out to us at hello@evolution.ai.
