Zillow ran an active institutional buying program earlier this decade, bidding for and purchasing scores of properties based on what its algorithmic models suggested was a fair value. But following a market price shock, the fortunes of the iBuying unit went south, and the company was forced to shutter the business.
The high-profile
Yet even though it's been a cause of prominent business snafus this decade, the concept of AI drift still poses challenges many struggle to fully understand, especially in data-dependent industries like mortgage.
"In the most layman's terms, I would define AI drift as the lack of training of AI data modeling," said Jason Bressler, chief technology officer at
Immediate outcomes include the delivery of false and inaccurate information, but those results can easily snowball for a mortgage lender when drift leads to noncompliance or approvals of unqualified borrowers. While effects can be serious, consistent monitoring helps mitigate ever-present risks, which pose an ongoing threat as generative and
Without robust due diligence that occurs regularly, "whatever you're getting probably has some level of drift to it. It's just a matter of how much," Bressler said.
In 2022, scholars at Harvard University, Massachusetts Institute of Technology and the University of Monterrey found some degree of drift occurring in 91% of deployed machine-learning models. With the rapid technological advancements in AI over the ensuing four years, generative tasks still experienced drift between 76% and 89% of the time, according to research conducted by European computer science experts in 2026.
What causes drift?
Any AI tool or large language model reliant on data is susceptible.
"It's always existed as long as we've been doing machine learning or any sort of AI development," said Brian Woodring, chief information officer at national
Where much of the mortgage industry might first encounter drift is in output coming from a rapidly expanding number of consumer-facing agentic chatbots and voice agents that lenders and servicers are introducing, as well as among the copilot assistants employed in business-to-business interactions.
Models can only be as good as the information contained within them, technology leaders regularly emphasize. A lapse in updating data or regulatory guidelines within an individual AI model can easily throw it out of alignment and produce flawed results.
"As the baseline assumptions change, and the data changes, the model will drift over time even if the algorithm that the model is built on never changes at all. The underlying data will change over time, and you will start to see models perform differently," Woodring explained.
Understanding where an AI tool obtains its knowledge is as important in discovering the causes and circumventing drift, tech leaders add. The inputs may not have originated from a human employee.
"If your AI is auto learning, then you can inadvertently introduce drift," said Justin DiPietro, co-founder and chief strategy officer at Glia, a fintech company that provides artificial intelligence technology to banks and credit unions.
"When you start looking at some of the new ways we're investing in having AI self-learn, that's where you can also have drift set in," he added.
How drift contributed to a business shutdown
When using flawed analysis based on drifting models in a lending capacity, the risks for a mortgage company run the gamut from bad lending decisions to
"Ultimately the model can't be held accountable," Woodring said. "If they make a terrible mistake, the model is not going to be the one responsible."
Lax oversight of the data businesses rely on is cited as a cause of multiple notable corporate fiascos this decade. Along with Zillow, Instacart also failed to quickly and adequately gauge how massive, sudden shifts in demand might skew its models in 2020 upon the onset of the Covid-19 pandemic. By not factoring in the effect of pandemic shock, the algorithmic models of both companies produced flawed results, leading to public relations disasters, and for Instacart the need for manual overrides.
In the case of Zillow Offers, the business platform neglected to account for the sudden slowdown in price growth after a massive Covid-19 spike, leading to an oversupply of properties the company found impossible to turn over quickly enough to gain profits. Given the uncertainty around home prices and interest rates, a similar concern could easily arise.
Pay attention to tone
Beyond data inaccuracy, new agentic AI tools also run the risk of harming a company's approval ratings as a result of interactions that may leave customers frustrated if escalated.
"When the model is challenged, and someone disagrees with it, how does it respond? How does it respond to a customer who's trying to divert the topic of conversation to something else? There are so many little nuances," Woodring said.
A preemptive strategy sets up conditions to reduce what's called tone drift and combat such shifts, technology leaders say. Key to the strategy are boundaries that restrict an AI chatbot or agent to discussing a limited set of specified topics before the tool is introduced publicly.
"We include guardrails in terms of acceptable content for the conversation," said Kevin Foley, director of product management at Optimal Blue, the secondary market software and data provider, whose products include an AI-backed virtual economist tool to assist lending stakeholders.
"That keeps the focus on the conversation and the professionalism within the conversation," Foley noted.
When a situation becomes heated or overly complicated, a human employee should always be on hand to take the reins, according to Woodring.
At the same time, development of AI tone and the ability to fine-tune it and add personality to tools presents a new marketing opportunity companies can also take advantage of to demonstrate expertise and shape public perception.
"The way we look at it is that you would give them practically everything about your brand that you can. Customers shouldn't feel like they are talking to OpenAI," said Instamortgage CEO Shashank Shekhar.
"You're not just saying, 'Talk to borrowers in a tone that's courteous,'" Shekhar continued. "You're talking about the need to be high touch. Are you educational? Are you a luxury brand? Are you veteran focused? That's a personalization of the tone."
How companies can manage drift
Because artificial intelligence functions at a highly advanced level and possesses the ability to make autonomous choices and vary responses to a query, its upkeep and maintenance demand attention and care beyond what many in the mortgage industry are accustomed to when dealing with traditional tech stacks.
"It is a new and different type of software that needs its own performance framework," Foley said.
An AI model is more akin to a human team member than a piece of technology and should be trained and managed in much the same way, according to executives.
"In many ways, it behaves like one, and it is a little bit unpredictable the way people often are," Woodring said. Instead, companies often fall into the trap of building an AI model and think their work is done, save for basic maintenance, he added.
Since they constantly absorb data, today's models need regular check-ins, much like workers who have earned promotions and gained new responsibilities.
"You trust them, but you do have to make sure that they're actually doing their new work correctly," Bressler said.
Depending on what the model is expected to produce, proper maintenance calls for frequent monitoring of responses, and even oversight from individual data scientists dedicated to each model, as is the case at UWM. Smaller companies relying on outside vendors for their AI need to make sure they ask the right questions and find out how often models are updated and how well providers understand mortgage operations, Bressler said.
Constant testing of output against actual historical results and internal data also helps ensure an AI tool doesn't deliver suboptimal results.
"You can measure it, and the way you would is by establishing certain performance benchmarks that you would expect to meet. It's all context dependent," Foley said.
"Part of that is making sure that they're trained on the most recent data, and they're incorporating the most recent inference data."
For regulated industries like mortgage or tax, disclosures of future policy changes mean companies can start managing any anticipated drift before regulations are mandated.
"We know the drift is coming, so we can predict it," DiPietro said, noting that new laws often spell out clear rollout dates.
Apart from comparing results against prior outcomes, a strategy companies can also use is testing AI models against others that have already proven their ability to deliver quality output.
Although intensive, AI model oversight, in the end, will serve a company's best interest, paying dividends as its data changes.
"When I think of model governance and maintaining a model, it's really about ensuring that it's to your advantage that the data changes over time," Woodring said. "You're making sure that the model is provided with the right context, with current context, that it's tested and evaluated so that it's getting better over time instead of worse."
