.png)
For the past 10+ years, finance-oriented data scientists have been perfecting machine learning using artificial intelligence. Many of these financial data scientists were trained in places like American Express and Capital One, where they used huge datasets of consumer data to determine the creditworthiness of prospective borrowers.
It started with traditional logistic regression analysis, where they would regress one, two, or three variables against a dataset of variables to determine which variables were most statistically significant when determining the likelihood of a default of a prospective borrower. They would run these regressions over and over again, regressing against hundreds of different combinations of variables. The variables were things like home ownership, car ownership, the number of credit cards outstanding, the total debt-to-income, and FICO score. As a result, they would find the mix of variables that were the best predictors of creditworthiness and would extend credit to prospective borrowers that fit their predictive models.
When machine learning emerged, the game was changed and data scientists upgraded from regressing a couple of variables to regressing all variables against all other variables in every combination possible, oftentimes resulting in tens or hundreds of millions of combinations. They used deep learning techniques to train neural networks with multiple layers of data so that they could understand the relationship between all variables. This type of analysis was beyond the capacity of humans and required machines to complete the computation. It was the dawn of machine learning.
In 2011, I co-founded a credit decisioning business called LendingRobot (originally called Lend Academy Investments), which applied logistic regression and neural networking to datasets from P2P lenders like LendingClub, Prosper, and a variety of additional originators. We figured out ways to lend to high-quality borrowers using cutting-edge data science and we were part of a cluster of innovators that built businesses based in this field of financial data science. We hired data scientists from the big firms, we innovated in ways to analyze the data, and we created new credit models that were more predictive than the legacy models. Eventually, those big firms, which were sources of talent, started adopting some of our methodologies. At this point, machine learning is mature and widely used across the industry in this capacity.
Machine learning has yielded some amazing results in the field of credit decisioning. As a result, lenders have been able to expand their set of qualified borrowers while minimizing default rates. In other words, lenders can lend to borrowers who may have otherwise been rejected using traditional analysis. This ability to minimize defaults and expand borrowing capacity to marginal borrowers has provided much-needed access to capital to people in need.
In this example, machine learning was used to make forward-looking estimations based on historical data sets. It analyzed structured data and predicted based on quantitative data. This was early financial AI and my first experience using this technology.
Machines have mastered human language and this will transform industries
We are now entering into a new phase of AI, which is incredibly exciting and has far larger implications for society. In this new era of Generative AI machines have mastered human language. It is about organizing unstructured data, analyzing, and predicting based on qualitative data. While the technology is not perfect, it is very impressive at this early stage, and it will continue to get smarter over time, likely at a very fast pace. Already, it has an amazing ability to read, write, speak, and reason and it can adjust for tone, inflection, and personality.
After the invention of transformer technology in 2017, which is the underlying neural network that powers large language models like OpenAI’s GPT-4 Turbo, the ability to process huge volumes of qualitative data like text, images and videos became possible.
Transformer technology utilizes an AI concept called attention to emphasize the weight of related words, which provides context for a word or token describing another type of data (image, video, etc). Data scientists have utilized this technology to compute language. Machines are now able to convert the meaning and context of words into data that is weighted and associated with other data. As a result, AI can produce the statistical probability of the next word in a sentence. In addition, AI can generate new word clusters. It came as somewhat of a surprise that transformer technology works amazingly well. When ChatGPT launched in late 2022, it demonstrated to the world that Generative AI is natural, intuitive, and intelligent.
“Generative AI is as revolutionary as mobile phones and the Internet”
- Bill Gates
Over the past 40 years, we have experienced a number of technological breakthroughs that have propelled technology forward and incorporated it deeper into our lives. Everything from the microprocessor to the graphical user interface (GUI) to the browser-based Internet. More recently, the latest paradigm shift was the introduction of mobile technology, which put computers into our pockets and made them an invaluable part of our everyday lives.
Generative AI is the latest breakthrough technology. We have all experienced Siri, Alexa, and Google Assistant so we have had a chance to interact with the early primitives in this field, but ChatGPT has demonstrated the major next step. Ask it anything and it will provide a comprehensive response that is easy to understand. While it is certainly not perfect, we can all imagine how much better Siri would be if it provided ChatGPT-like responses, which will happen in a short matter of time.
Over the next few years, there is going to be a huge wave of new natively built Generative AI applications that will unlock the power of Generative AI in ways we haven’t dreamed of yet.
My co-founders and I have committed ourselves to reinventing accounting using Generative AI. We want to build a native Generative AI application that was not possible before this new paradigm was introduced. In order to do so, we have to understand the groundwork of what to focus on and what to avoid.
The first thing to understand is the underlying infrastructure. Whenever you build on another layer of technology you run the risk of unintended consequences based on changes to the underlying technology. The key components of the Generative AI tech stack start with LLMs. We chunk text and parse data into an embedding model that is stored in a vector database. The data is hosted in a cloud service and in instances where computation is involved, we have the LLMs to generate SQL responses to run through Python to generate computational responses.
What we realized is that LLMs are generalists and there are numerous ways to improve the quality of a response to a query. Tellen has the opportunity to build a solution that utilizes this underlying infrastructure to generate high quality accounting-related responses.
Given our deep experience building fintech applications, we have chosen financial services as our vertical. But financial services is a huge industry and we want a specific use case that is ideal for Generative AI. We started thinking about what work functions require knowledge workers who organize, analyze, and predict based on large volumes of quantitative and qualitative data. We landed on accounting and specifically financial audits as a great use case for Generative AI.
Here are some of the facts related to the US financial audit industry:
We have conducted more than 90 product research calls with accounting and accounting-related professionals and our takeaway is that Generative AI is a high priority for accounting firms and their first step is often to launch a private instance of an LLM that is secure and confidential that can handle their firm and client data. As a first step, Tellen will provide Gen AI Accounting Infrastructure to accounting firms so that they can launch their own Chatbot. With this foundation in place, Tellen will begin to build AI audit applications.
When an auditor renders a judgment on the validity of a set of financial statements they utilize three sets of data in order to make their judgment: publicly available data (GAAP, IFRS, Tax Code, etc), firmwide methodologies (a typical Big 4 audit methodology handbook is about 7,000 pages), and client data, which must be validated. Audits take tens of thousands of hours to complete and require millions of dollars because they require meticulous analysis and they must reconcile huge amounts of unstructured data (ie, bank statements to invoices). Auditors must make sure the companies have produced proper disclosures and risk standards. They must make sure that the numbers are referenced and underlying documents are cited as evidence.
The company pays the auditor based on time spent on the audit. The auditor’s reputation dictates the price that they can charge for an audit. By securing a clean audit by an audit firm that has a strong reputation, an audit becomes a cornerstone of trust in our capital markets system.
Tellen’s audit management solution generates contextualized accounting related answers by optimizing LLMs through prompt engineering, retrieval-augmented generation (RAG), and fine-tuning. As a result, our solution will generate answers that are specifically tailored to our industry generating higher quality responses than if the same questions were asked directly into an LLM.
We will continue to refine and roll out Generative AI tools as part of our long-term roadmap to build a full-suite AI-powered audit management system. Future clients will be able to add their own proprietary data, including firm handbooks and methodologies, to generate responses that are specific to their needs.
We believe that our solution will significantly increase the productivity of auditors, increase the quality of the audit, and pass along margin improvements to accounting firms and their clients. Auditors will be able to take on more clients and will render judgment based on better data. We think that accounting firms will accelerate revenue AND expand EBITDA margins by 25% or more. In an era where accountants are in short supply and the quality of audits is increasingly deficient, we think that our solution cannot arrive at a better time.
We will start with the release of Tellen’s accounting chatbot and we encourage anyone in the accounting industry to use it and share it if you think it is helpful. We will be refining the chatbot based on your feedback so let us know what you think. We will use our learnings from this free product to inform the development of our enterprise audit management system.
.svg)
.svg)
