We are thrilled to announce that Informed uses the OpenAI models that power ChatGPT through Microsoft Azure. This further enhances our ability to deliver the very best to our clients, empowering them to stay ahead of the technology curve.
For years, Informed has used machine learning and AI to help lenders streamline their lending process, lower the cost of credit, and ensure that loan applications are fairly evaluated; all while making loans more accessible. Over the past few months, large language models (LLMs), such as ChatGPT, have taken the world by storm and opened new techniques that Informed uses to make loan processing even more effective.
Large language models are part of a different class of AI models called foundation models. They contrast with application-specific models trained on task-specific data to perform a specific function. ChatGPT is trained on planet-scale datasets. What gives foundation models their superpower is the ability to perform many different tasks and functions. This is accomplished simply by prompting (no retraining required) or tuning the model with a small amount of data. However, hosting foundation models requires a large amount of compute power, and they are vulnerable to producing inaccurate results. They suffer from “hallucination.” This is when an AI model produces a confident response that’s not justified by the underlying training data.
“Informed is the consumer lending industry’s only vendor mitigating AI hallucinations by performing Few-Shot Learning using both credit application and approval data along with historical insights from our contributory database consisting of over 50 million records.“
In doing so, Informed eliminated 98% of hallucinations that result in factual errors and false positives affecting those who apply AI to optical character recognition (OCR) output from documents. Informed’s additional checks and controls lead to best-in-class accuracy and real-time decisioning for higher capture/conversion rates, lower operational costs, and improved compliance.
We strive to be the ultimate technology partner equipping clients with cutting-edge technology. And we already provide best-in-class document classification, extraction, and an AI-enhanced rules engine, among other innovations.
Informed strictly governs how data is processed, used, and stored. We partnered with Microsoft Azure instead of working directly with OpenAI. This enabled us to implement specific controls around the retention of data and limit use. To learn more about Microsoft’s variations of OpenAI models, visit: https://learn.microsoft.com/en-us/legal/cognitive-services/openai/data-privacy.
We implore others to leverage the Microsoft Azure OpenAI models further restricting access to sensitive data by applying for modified abuse monitoring via https://aka.ms/oai/modifiedaccess. And, we continue to recommend industry best practices as financial institutions and lenders seek to leverage the latest in Artificial Intelligence.
Below we share how we harness this groundbreaking technology. While exploring the vast potential, our applied machine learning engineers see both opportunities and challenges. Since ChatGPT can provide inaccurate outputs, we have committed to thorough statistical testing as per our written Model Validation and Governance policies before introducing it to our industry partners. We are committed to an AI-enhanced lending process that remains seamless, secure, and accurate.
In this multi-part blog series we will explore different use cases with increasing amounts of data input to ChatGPT, anonymizing wherever necessary.
Let’s look at two examples of using the ChatGPT with “group of words or paragraph” as the input. We will present actual transcripts so that you can see the technology in action.
Employer Name Matching
ChatGPT has trained on vast amounts of information available on the internet and delivers astounding results. At Informed, we are refining the process of comparing two employer names, a critical task in employment verification. For instance, you may work at Audible, but your paystub and/or bank statement might say Amazon (Audible’s parent). We have made great strides in addressing this challenge and achieved remarkable outcomes.
Now we are taking our capabilities to the next level by integrating ChatGPT technology into our system. Check out the image below to see some impressive results on the employer name match challenge. And this is just the beginning of what we will accomplish.
We won’t stop here – these enhanced capabilities provide deeper insights into organizations. This added layer of understanding helps identify potentially fictitious companies or those linked to fraudulent activities. We constantly improve our technology ensuring our clients have a secure and reliable lending process.
Understanding Late Fees
Contracts and ancillary documents often contain complex language that needs careful interpretation. This includes legal requirements, understanding cancellation policies, or deciphering late fees. Let’s take a look at how ChatGPT helps make sense of it all.
For example, a borrower may receive this notice: “If we do not receive your entire payment within 15 days after it is due (10 days if you are buying a heavy commercial vehicle), you will pay a late charge of 5% of the scheduled payment or $50, whichever is greater.”
To understand the late fees, an individual has to go through several steps. They need to calculate the cut-off date for the late payment each month. Then, note the type of vehicle and compute 5% of the payment. Finally they need to know which is higher – the percentage or the flat fee. Depending on circumstances, they might also have to check the state laws to ensure the fee doesn’t exceed the cap.
Future installments in this series will focus on other use cases for advanced AI in loan processing. In this initial post, we explored two use cases based on different types of data fed into ChatGPT. Foundation models like ChatGPT seem superhuman, because they almost instantly produce useful results for a variety of problems. Yet, we must proceed with caution. The technology is in its infancy and the outputs can be inaccurate. The high levels of accuracy needed in loan processing necessitates that we adopt this technology with care and only after rigorous testing. Watch this space for part 2 on this topic.
In the meantime, contact us to see how our technology can improve your lending process, lead to higher capture rates, lower operating expenses, and improved compliance.
Nishit Kumar is Head of Machine Learning at Informed and has been an engineering leader in tech startups for over 20 years. He is developer of AI products with expertise in machine learning, deep learning, computer vision, and AI.
Jatin Agrawal is a product lead at Informed. With over five years of experience in Deep Learning and a background at AI startups and Microsoft, he is committed to improving auto lending with innovative AI products. His work is published in numerous publications.