Three distinct questions, not one
To address the complexities of AI-to-AI communication, it is helpful to deconstruct the broad concept of "AI agent trust" into three specific, manageable questions.
The first is disclosure. When an AI agent calls a servicer on behalf of a customer, does it have a legal and ethical obligation to state clearly that it is an artificial intelligence? There is a strong argument that the answer must be yes. We already have established rules requiring the disclosure of call recording or the use of automated dialing systems. Transparency is the bedrock of financial interactions; a servicer must know whether they are negotiating with a person or an algorithm to ensure the conversation is handled through the appropriate protocols.
The second is authentication. Even if an agent identifies itself honestly, how does the servicer know that this specific agent was actually authorized by the customer? The payments industry is already working on technical solutions for this. Visa’s "Intelligent Commerce" and Mastercard’s "Agent Pay" are frameworks that allow AI agents to prove their authorization and operate within strictly defined spending limits. However, no such standardized framework currently exists for the more complex world of loan servicing, where the stakes involve long-term debt and home ownership rather than simple retail transactions.
The third question, and perhaps the most overlooked, is ongoing fitness. This is where my own experience with model drift becomes highly relevant. Even if an agent is properly authenticated today, is it still capable of representing the customer’s interests effectively six months from now? AI models are not static; they change based on new data and environment shifts. An agent authorized last year may have quietly degraded in quality, yet most current credentialing frameworks only focus on the initial "handshake" of authorization. They remain silent on the question of whether the agent is still "fit" to perform its duties.
Where this leaves the asymmetry question
A common counterargument in the tech community is one of symmetry: If a mortgage servicer is allowed to use its own AI agent to speak with customers, why shouldn’t the customer be allowed to use an AI agent to speak back on equal terms?
The answer to this lies in accountability, not in a fear of the technology itself. When a mortgage servicer deploys an AI agent, there is a clear line of responsibility. The servicer owns the system, manages the monitoring infrastructure, and carries the legal liability if the system fails or provides incorrect information. In a corporate environment, a system cannot "own" an outcome; only a person can. The servicer has designated individuals who are ultimately responsible for the agent’s actions.
Most individual consumers do not have this level of governance. If a consumer’s AI agent fails, the consumer bears the cost, but the servicer is left dealing with the fallout of an unreliable negotiation or an unauthorized request. There is a fundamental difference between "personal acceleration"—using AI to make life a little easier—and "mission-critical automation."
When an AI agent is used to negotiate a loan modification or dispute a charge, it is performing a mission-critical task. Treating a consumer’s unmonitored AI agent as the equivalent of a highly governed corporate AI system ignores the inherent risks and the lack of accountability on the consumer side. A recent Forrester survey highlighted this gap in trust, finding that only about 24% of U.S. online adults trust an AI agent to act on their behalf even for routine purchases, let alone complex financial restructuring.
