How in-house AI coding risks unexpected token, upkeep costs

Mortgage lenders tempted to let AI ‘vibe code’ their way to more affordable in-house software may be underestimating the price tag and the risk. As tools like Claude and ChatGPT make it possible for non-developers to build tech in a matter of prompts, industry advisors warn that token costs, compliance exposure and staff turnover can quickly erase any savings over buying from a vendor and more than 40% of agentic AI projects are on pace to be abandoned by 2027.

Processing Content

While tech investment strategies are unique to every business and depend on a host of factors, including internal expertise, appetite for risk and experimentation and — maybe most importantly — budget, the ease of using AI means some companies are taking a look at ways to create their own tools for specific tasks. 

Register HERE to learn more about BUILD V. BUY at the  DIGITAL MORTGAGE CONFERENCE

“It’s as easy as ever, even for folks that have very little development experience to delve into that world,” said Devin Zito, director of information services and corporate counsel at national mortgage lender Assurance Financial. Tools like Anthropic’s Claude and OpenAI’s ChatGPT are creating an environment to support such efforts, he said. 

The vibe code surprise

Using platforms like Claude or ChatGPT to create mortgage processing tools, though, carries costs that are catching some companies by surprise. The AI giants charge for their artificial intelligence development services based on metered use, with tokens that come with subscriptions or prepaid plans serving as the currency. 

With costs of tasks typically not predetermined, a fair number of projects are abandoned as businesses realize how many tokens are being consumed. More than 40% of agentic AI projects are likely to be abandoned by the end of 2027, with unexpected higher costs and uncertain return on investment cited as primary reasons, according to a 2025 study from technology research firm Gartner. 

Even with the quantity of tokens needed, though, building certain types of technology can still hold value in some cases, when it adds capabilities a company may have not previously thought it would be able to produce.

“There is cost, but certainly not anything like hiring a full-time developer or a full-time development team to go off and do the same things. But the technology pros have to have the knowledge to be able to direct it appropriately and to get the results that they want,” Zito said.

With AI still a relatively new concept, users can expect to see token expenses come down as well, just as they have historically for other technology as production improved over time, suggested Erik Eggers, chief revenue officer at Rocktop Technologies, the mortgage solutions and data services advisory firm.

“As users are learning how to control their modeling, I think the providers are also working on ways to make their models more efficient so that the token cost is more reasonable for the end users,” he said.

While the AI platforms are opening the door to internal tech development, the same regulatory compliance rules apply, an aspect that might be easy to forget with their ease of use and the initial excitement of creating a tool. 

“I don’t know if that changes very many of the concerns with respect to if you build it versus buy it,” said Zito, who also serves as a business and technology attorney in Baton Rouge, Louisiana.

The rules of maintenance also remain the same, said Eggers.

“AI tools and vibe coding have collapsed the cost of building software on the front end, but what it didn’t do was collapse the cost of everything around the tools,” he added.

With vibe coding, a mortgage team can quickly develop highly functional software, such as a copilot, on their own to deliver up-to-date information, guidelines or calculations in just a series of user prompts. But AI’s rapid improvement now allows for much more, making it possible for companies to take volumes of unstructured data from a single origination file or full portfolio of loans and quickly perform complex analysis that guides their decision-making.

The capability of AI today is light years ahead of what older technologies many in the mortgage industry used to rely on, like optical character recognition, could do thanks to current digital advancements and integrations. 

“Outside of the prompts, you’re using other tools that are available now, where it has access to your system register of files and an unlimited reservoir of memory that it can hold for you,” Eggers said.

The cost of oversight

After the initial phases of development, maintenance of the tools created may bring other unexpected challenges down the road for businesses that don’t plan ahead. High among them is the regular churn of personnel, including tech executives, and the accompanying loss of institutional knowledge whenever that occurs.   

“People are always coming and going. The mortgage industry is kind of like musical chairs,” said John Geertsema, managing principal at financial services advisory firm Capco. “There’s a lot less risk hanging your hat on a vendor solution.”

Yet having the support of specialists at a third-party vendor does not remove the onus on their mortgage clients to maintain proper upkeep and ensure they have internal expertise available for a purchased tool. The effort needed for compliance might be as rigorous for a vendor-bought solution, especially as the government-sponsored enterprises are now emphasizing to mortgage companies their responsibility to understand everything about their AI technology, just as they do for other aspects of their business.

“If we have a vendor that’s employing AI in the course of solving whatever we’re asking them to solve for us, those due diligence components have to happen,” said Zito, who also serves as a business and technology attorney in Baton Rouge, Louisiana. “If we build for ourselves, that doesn’t go away. We just become that vendor that has to provide that due diligence for ourselves.”

There is also no guarantee that third-party products will keep up with the pace that AI is improving with timely updates as needed, according to InstaMortgage CEO Shashank Shekhar. 

“We don’t know if they will continue to evolve as quickly as the AI evolves, unless your contract says that they will,” he said about purchased technology. “There too, it’s very hard to prove they are constantly evolving,” Shekhar remarked. 

Ultimately, the right answer at a company for whether to build, buy or do a combination of both may lie in how unique the tech development is and how well it can serve the needs of its own employees, Geertsema suggested. 

“If it’s a commodity, like something that already exists in the market, many products provide this capability very well,” he said. “When you want to build is if you have some type of competitive differentiation that is like a secret sauce.”