By Blake Gibson
If you're a credit union leader, you may be exploring options for artificial intelligence (AI) for loan origination and thinking, “That sounds like something we could use to boost efficiency; let’s build it in-house.”
That would be an unwise leap of faith. Implementing AI for loan origination is far from a quick or simple task. In fact, it’s a sophisticated and expensive endeavor. And there’s a lot of risk involved: AI regulation was listed as a “risk factor” in the annual reports of 137 Fortune 500 companies.
This is a field that demands strategic planning, significant investment and specialized expertise. Success with AI in lending takes years of careful preparation, data testing, model training and consistent financial commitment.
How Much Commitment Are We Talking?
You’re probably familiar with ChatGPT, the popular language model created by OpenAI. ChatGPT-4, for instance, was trained using trillions of words and thousands of high-powered computers. The estimated cost for developing that model? Over $100 million.
Looking ahead, the upcoming ChatGPT-5 is projected to cost between $1.7 and $2.5 billion to develop—17.5 times more expensive than its predecessor and nearly 400 times more costly than ChatGPT-3.
These costs are driven by high-demand computing needs, increasing GPU prices, and expensive cloud infrastructure. Add to that the fact that AI research and development requires top-tier data scientists, whose expertise doesn’t come cheap.
The Hidden Costs: Accuracy and Integrity
The expense doesn’t stop there. The true cost of AI goes beyond just training and infrastructure. There’s the often-overlooked “cost of truth.”
It takes a massive amount of validation to ensure an AI model is accurate. Think about it—if we’re building an AI model that needs to verify “truths” across trillions of data points, we would need thousands of workers just to validate this information. Even if you could hire people at $5 an hour (an unlikely scenario), this validation effort would cost billions of dollars. It's simply not feasible to match that scale using human labor.
Moreover, this process would require an unheard-of level of consistency across a massive workforce—something that’s inherently difficult when humans interpret data differently. This illustrates why AI often generates outputs based on probabilities rather than precise accuracy, focusing on “the most likely next token” rather than “the most accurate next token.”
The Challenge of Model Drift
AI models also face what’s known as "concept drift," where data and trends evolve over time. In the mortgage world, this can present significant challenges. Take tax forms, for example—they change only slightly from year to year. In contrast, other documents, like pay stubs, are more likely to undergo rapid changes as new templates and data formats emerge. Similarly, bank statements can differ dramatically from one financial institution to another, which makes modeling them accurately an ongoing challenge.
As data changes, the AI must adapt quickly, and that requires constant tuning and monitoring. If this isn’t done properly, the AI model may produce inaccurate or unreliable results, which could lead to costly mistakes.
It’s Worth Partnering with Experts
Clearly, developing AI for loan origination is no small feat. It’s not something credit unions can afford to take lightly. It requires a partner with the expertise, infrastructure and resources to manage such a complex and costly project.
That’s where companies like Dark Matter Technologies come in. With years of experience in AI and a deep understanding of the mortgage industry, we have developed a sophisticated suite of AI-driven tools designed specifically for intelligent document processing. Our suite of AIVA® AI Virtual Assistants work tirelessly to improve document accuracy, streamline workflows and enhance automation—all critical to mortgage origination.
Dark Matter has invested tens of millions of dollars into developing our AI technology, and we employ a team of experienced data scientists, including several Ph.D.s, who’ve been working with AI for decades. This is not something a credit union can replicate overnight.
In the past three years alone, we’ve processed over 356 million pages and 105 million documents. Our AI has extracted over 148 million data labels, delivering consistently high performance. In fact, our AIVA Docs solution now supports over 1,100 document types and 1,500 data elements with extremely high accuracy. Plus, it integrates seamlessly with leading loan origination systems, including the Empower® LOS platform.
The Risky Side of D-I-Y
Now, imagine trying to develop this technology in-house. Without the proper resources, training and ongoing investment, it’s easy for models to become inaccurate, biased or inefficient. If a self-built model isn’t rigorously tested and properly calibrated, it could jeopardize your business’s success and put you at risk of compliance issues, delayed loan processes or even financial loss.
Are you still convinced that D-I-Y AI is the right choice for your credit union?
True, AI in loan origination is a game-changer, but it’s a complex one. It pays to partner with the right experts—companies that have invested the time, resources and expertise to make AI work at scale. When you work with the right AI provider, you can focus on what matters most: providing members with fast, reliable and friendly lending experiences.
A seasoned and tested attorney and product leader with experience creating and leading legal and business teams, Blake Gibson is currently general counsel and chief operating officer at Dark Matter Technologies.