Mortgage AI goes live, humans stay in loop

By late summer, the strongest evidence was no longer slideware but live rollout. In Lendi Group’s August case study, the company said it was “moving beyond pilots to production tools designed for its 1,300 brokers,” with chief product officer Travis Tyler adding that “over 70% of our development now is on AI products”; just as important, the rollout was governed in production, with Tyler saying, “Where there is anything that is regulatory or needs to be 100%, it's deterministic,” while AI could “present options before handing the process to a broker,” and Lendi planned to widen access and study day-to-day use.

Other deployments in the same window showed the same pattern: real-world testing paired with explicit controls. Lendi’s policy agent “consists of a manager and 28 agents” using “about 10,000 pages of policy material,” and after testing showed brokers wanted proof, it added links to source documents; “Twenty brokers helped build and test the agents.” Customer Experience Dive reported Mutual of Omaha Mortgage had already put borrowers and employees “into the virtual hands of voice AI agents,” but only for the first 30 seconds before handoff, while Figure said it has “over 500 partners” and uses its own production originations “as a training ground.”

The operating logic is not AI replacing mortgage labor, but AI taking the repeatable share while humans retain the judgment layer that keeps decisions defensible. As Neil Sahota put it in Chrisman Commentary, “How can we try and get from 94 to 99.9%? We need humans,” and that principle now shows up in production workflows where AI is “handling inbound and outbound borrower communications, reviewing loan files, identifying missing documentation, and automating as much as 80% of repetitive underwriting tasks,” while staff focus on the exceptions, policy fit and borrower-facing decisions machines should not own alone. That same division of labor is extending to the front end: “We’re seeing voice AI being used both for managing inbound calls and for like outbound calls to call leads… In the case of one large originator, we were able to 4x the number of appointments that were being set,” but the model still depends on people to own the consequential judgment and borrower trust.

What makes the model scalable is that the handoff point is measurable: in one case, “The AI is now able to automate 80% of that work,” so “their team is able to spend time on the 20% that actually requires human judgment, is complicated, cannot be automated away.” That split cut review time from “about 4 hours” spent on income calculations and agency-rule comparisons to “under an hour,” while another lender said staff can write credit memos in 20 minutes instead of three hours and “scale up” without needing to “add 50 employees”; at bank scale, Westpac says its AI agents process more than 32,000 payslips and 1.5 million transactions each week.

The fight in mortgage AI has narrowed to a trust boundary: not whether automation helps, but when its output is reliable enough that a human no longer has to check it. In Product Talk, one speaker framed the issue as expectation-setting for a system that still “need[s] to be trained,” adding, “I don't see we're going to achieve 100% accuracy with AI… You can't be 100%. You can be maybe 99.95, but you can't be 100%,” which is exactly why the autonomy debate persists.

Skeptics argue that once lenders treat near-accurate AI as fully trustworthy, the downside is not just technical error but borrower harm and regulatory exposure. Chrisman Commentary noted that “55% of Americans still want a human involved in their mortgage… up 14% from last year,” while The Best Ever CRE Show captured the operational version of the same concern: “you’ve got speed but not necessarily 100% trustable accuracy,” so “it’s not enough to build just another wrapper on top of a model,” especially when a governance layer sets “a threshold of 90%” certainty for escalation.