The landscape of mortgage servicing and consumer financial services is on the precipice of a fundamental shift. While it is not yet a daily occurrence for a mortgage servicer to receive a call from a customer’s autonomous AI agent, the technological and regulatory infrastructure required to make this a reality is converging with remarkable speed. As artificial intelligence moves from being a tool for human assistance to an "agentic" entity capable of making decisions and initiating actions, the financial sector faces a new set of challenges regarding trust, accountability, and security.
According to a recent International Monetary Fund (IMF) staff note, agentic AI is already beginning to influence the world of payments, steering the industry toward a future defined by agent-mediated decisions. This trend is further supported by a 2026 Cloud Security Alliance survey, which revealed that 85% of respondents in the financial services sector expect AI agents to begin initiating payments on behalf of consumers in the near future. For mortgage servicers, this means the day is fast approaching when a borrower’s AI might call to negotiate a repayment plan, dispute a fee, or request a loan modification.
Organizations that wait until these interactions become routine before developing a response strategy will find themselves in a precarious position. Writing policy under the pressure of a crisis—after a system failure or a security breach—is rarely effective. The institutions that take the time to think through these implications now will be the ones to set the industry standards, rather than being forced to react to the standards established by others.
How I got here
The urgency of this issue became clear to me following a specific incident involving a customer-facing AI system for which I was responsible. Over the course of several weeks, the system’s performance began to quietly and subtly degrade. It wasn’t a catastrophic failure that triggered immediate alarms; instead, it was a slow erosion of quality that went unnoticed by human supervisors for some time.
This experience led to an uncomfortable but necessary realization: If an organization’s own AI agent—a system built, managed, and monitored by internal experts—can lose its efficacy without anyone noticing, what level of confidence can we have in an agent built by a third party? When a customer’s AI agent calls a servicer, the servicer has zero visibility into that agent’s underlying logic, its training data, or its current state of "fitness." Relying on an external black box to represent a borrower’s best interests introduces a layer of risk that the industry is currently ill-equipped to handle.
Why the existing authentication playbook doesn’t transfer
The traditional method for verifying a caller’s identity in the mortgage industry is well-established. Servicers use a series of questions to confirm that the person on the line is either the borrower or a properly authorized representative. This process relies on "knowledge-based authentication," asking for information that only the real borrower would know. However, this entire framework is built on the assumption that the entity on the other end of the line is a human being.
The playbook fails to account for a more complex scenario: what happens when the entity correctly answering every verification question is not a person, but an AI agent acting on a person’s behalf?
The instinctive reaction from many in the industry is to simply apply the same verification standards. If the AI knows the Social Security number, the property address, and the last payment amount, it should be treated like any other authorized caller. But this approach misses a critical distinction. A human representative, once authorized and trained, does not typically undergo a "silent degradation" of their faculties over time. They do not suddenly lose the ability to understand context or start hallucinating facts. An AI agent, however, can suffer from model drift or software updates that fundamentally change its behavior and reliability. A framework designed to verify a human’s identity is insufficient when the caller is a machine that can "break" in ways a human cannot.
