Perhaps you’ve been reading up about AI for mortgage origination and thinking, “Yeah, we could use something like that to help our employees work faster. Let’s build it ourselves. Can’t we just plug in something off the shelf?”
I’m here to tell you harnessing the power of artificial intelligence (AI) for loan origination isn’t a breezy summer project or a simple plug-in. As a matter of fact, 137 of the Fortune 500 cited AI regulation as a risk factor in their annual reports. AI is a complex undertaking that enables machines to learn from data and recognize patterns to perform tasks more efficiently and effectively at a speed and scale that is revolutionary. Success with AI in the mortgage space takes years of planning and training, testing inputs, processing, outcomes, curated guardrails, adjustments, and assessments. And, of course, heavy and ongoing infusions of capital.
Companies Are Spending Millions for Research, Training
Most people are familiar with ChatGPT, developed by OpenAI. ChatGPT-4, a recent iteration, was trained using trillions of words of text and many thousands of powerful computer chips to answer your questions or rewrite passages. It’s estimated that OpenAI spent more than $100 million to train that single model.
The cost to train ChatGPT-5, expected in late 2024 or early 2025, is estimated at $1.7 to $2.5 billion. That would make it 17.5 times more expensive than ChatGPT-4 and nearly 400 times more expensive than ChatGPT-3. In addition to those rising costs, storage is not free, and compute time is expensive, especially for highly provisioned cloud systems. GPU costs are rising under heavy demand, and the hourly rate of data sciences to work these problems is some of the most expensive in an organization.
The Costs of Truth
Beyond the training and compute costs, there are additional layers of expense. The “cost of truth” is not often discussed. Consider for a moment the scale of a 1 trillion parameter model. While not an exact reference, for the sake of argument, let’s assume that this number represents 1 trillion “true things.” A person must generate that truth, and to ensure that the modeling feed is valid, a person should examine it to ensure validity (a process known as data labeling).
Assuming a human being could correctly identify one true thing every five seconds, and could keep that pace constantly 10 hours a day, 365 days a year, it would take 380,517 employees to validate 1 trillion truths. If you paid this workforce a globally adjusted wage of $5 an hour, that validation effort would cost just shy of $7 billion. The impossible constraints of pace and wage aside, it’s very clear that this scale is not viable in human capacity or economic terms.
By the way, in this example, none of those 380,517 employees could ever differ in their opinion. We all know the frustration we get from CAPTCHA boxes because opinions differ! This underscores why generative AI models focus more on “the most reasonable next token” versus “the most accurate next token.”
The Velocity Problem of Reinforcement Learning
The challenges around costs and scale are further compounded when velocity and model drift are considered. As the use cases for AI have moved out of purely technical data problems and into cases where models are taking on proxies for human behavior, such as documents’ representation of data, the data moves out of a state of immutable truth and into a state of higher velocity concept drift.
What is true today may not be true tomorrow, and the opinions and decisions of today may no longer be valid tomorrow.
For example, the thousands of document classes relevant to mortgage documents operate at different functional velocities. Federal tax forms change slowly, if at all, with minor header and form year updates. These documents, which are easier to granularly model, are less susceptible to concept drift. The more challenging pay stub class is much more likely to fall into drift as new and popularly emergent form templates are created or enterprise system generators such as Workday or ADP update product features to represent pay content in new ways. Also, consider the complexity of bank statements. Different financial institutions present common data (credits, debits, balances, etc.) in different formats and tables. Some may put the summation at the top of page 1, while others may put it at the end of a table on page 8. These nuances, which feel minor at first blush, create massive hurdles for data science teams to create models that are accurate.
It Pays to Partner with a Pro
There are far too many complexities in developing AI for mortgage origination to list in one blog. Suffice it to say, this is not a task for mortgage lenders to take lightly. You’ll want a partner well versed in AI, with dedicated resources developed over time and with deep pockets to continue to evolve their models.
Dark Matter Technologies has deep roots in the mortgage industry, and we began working with AI and machine learning years before the generative AI hype cycle began. Our goal is to reshape the mortgage industry with seamless, efficient and automated experiences while reducing the cost of originating a mortgage.
Utilizing AI and machine learning (ML) with the Dark Matter suite of AI virtual assistants (Aiva®), we have made great strides in improving accuracy and automation in intelligent document processing. This has led to faster, more automated, and, in some cases, extremely accurate performance around the complicated and nuanced set of mortgage origination documents.
We bring a unique blend of market position, experience, innovation, investment and performance at scale that is not easily matched by emerging providers pushing LLM API-based solutions or by organizations that have long been down a path of digitization first and discriminant rules approaches to content identification and extraction.
We’ve invested tens of millions in developing our Aiva product suite. Plus, we have dozens of long-tenured AI data scientists on staff, including several Ph.D.s who have worked with AI for decades.
In the last three years alone, Dark Matter has processed over 356 million pages across 105 million documents and has extracted over 148 million data labels through our Aiva Docs intelligent document processing solution. As of March 2024, Aiva is automating around 400 document types and 370 labels into the Empower® loan origination system at or above 99% F1 score accuracy. Aiva also supports hundreds of more documents at accuracy levels that meet or beat humans.
Now think of everything that could go wrong if you were to attempt doing it yourself. If your model isn’t properly trained or tested, shows bias or learns bad habits, you put your entire business at risk.
Do you still think this is a task you want to tackle?