Customers’ negative perceptions of chatbots and online chat widgets have been a barrier to mortgage companies’ successful escalations of AI deployment in customer service.

According to a new study from J.D. Power, those perceptions may be changing, opening a door for the mortgage space to improve customer satisfaction. However, it appears clear that there’s still a sizeable customer experience gap between interacting with an AI and chatting with a real human being.

J.D. Power’s Lending Intelligence Report noted that early iterations of online chatbots often left customers feeling like they were merely wasting their time. That feeling can lead to users opting for other means of seeking customer service: Of the 21% of mortgage servicing customers who experienced a problem in the past year, just 9% went through online chat as their first point of contact. Compare that to 48% who called customer service.

But preferences evolve over generations, and the same appears to be happening here. Customers from Generations Y (of which millennials are a subset) and Z are three times more likely to use online chat than older customers, according to J.D. Power’s data.

At the very least, customers who engage in customer service chats find them to be helpful. Sixty-seven percent of those who used a chat said they did so to try to resolve an issue; of those people, 83% said that the chat led to their problem getting resolved. Predictably, those who had their issue resolved via the chat had a higher customer satisfaction rating (702 points on a scale of 1,000) than those who didn’t (482 points).

Nevertheless, there remains a meaningful impact on a customer’s satisfaction when they feel they may be interacting with an AI compared to when they know they are working with a human being. Overall satisfaction dips 97 points, for example, when a customer is not sure they are interacting with a human. And the percentage of instances where the contact issue getting resolved drops 19% when customers are unsure there’s a human on the other end, per J.D. Power data. Perhaps most importantly, the share of users who report they will definitely reuse the same lender sinks by 20 percentage points if those users aren’t sure they’re working with a person or a machine.

This is a challenge for mortgage companies who are looking to AI as a way to cut customer service costs, J.D. Power’s report said.

“An investment in AI needs to represent a clear understanding of what the customer wants in terms of service and problem resolution, and how they interact with their customer service channels,” the report said. “Without that, customers may simply refuse to engage, leaving lenders on the hook for the time and resources spent on underutilized technology. Those who can thread this needle will see higher customer satisfaction scores, improved processes, and streamlined costs.”