A customer types one angry sentence into a chatbot: “My loan officer was rude and reversed my mortgage approval. Write a strong complaint and cite whatever will force an investigation.”
Seconds later, the lender receives a polished letter that may invoke formal mortgage-servicing rules, demand a documented review, and request preservation of records. The customer spent one sentence and almost no effort.
The lender cannot answer using one sentence. Someone must find the file, review the calls and emails, determine what happened, assess the organization’s obligations, and document the response.
The reply may be symmetric. The cost is not.
Walk into almost any lender’s boardroom and you will hear two questions: How do we do more with less using AI? And what capabilities should we build? Both matter. But together, they are still only half of the question.
While you have been building an agent to send out to customers, customers have been gaining agents to send back to you. Between the two sits another AI layer—a chatbot, copilot, search experience, or consumer application—that can filter, reframe, and interpret the exchange before a human is involved.
You are optimizing the agent you control. The other two owe you nothing.
Lenders are no longer only interacting with customers. Their own governed agents may increasingly encounter customer-controlled agents, while an intermediary AI layer can shape how information is interpreted along the way. The models may be similar, but the operating conditions are not: the lender’s agent must be consistent and defensible, while external agents can optimize each interaction for leverage.
This shift is already underway. Gartner predicts that by 2028, 70% of customer service journeys will begin and be resolved through conversational, third-party assistants built into consumers’ mobile devices. That means more customer interactions may be mediated by AI interfaces the enterprise does not own or control. (Gartner, 2025)
External agents make three moves
Strip away the growing list of use cases and the pattern becomes simpler. External agents judge, manufacture, and optimize.
1. Judge: They shape the first impression
Before a prospective customer contacts you, they can ask an AI: “What are the main complaints about this company?”
In seconds, the AI may return a tidy list: slow underwriting, missed closing dates, inconsistent communication, surprise fees, hard-to-reach support, or denials with little explanation. Research that once required hours of searching now takes one prompt.
But retrieval is not the most important change. The AI can also conclude.
It may surface thousands of complaints without the far larger number of transactions that are closed without trouble. The count arrives without the denominator, and volume quietly becomes a verdict. Then the system may add a grade or recommendation: “I would lean toward a different lender.”
Your first impression has already been formed—in a conversation you did not join, by software that may never hear your side.
2. Manufacture: They create pressure on demand
The complaint in the opening is powerful because the effort required to produce a credible escalation has collapsed.
A customer’s agent can create a new complaint, appeal, chargeback, hardship request, or negotiation in seconds. A top Canadian mortgage company, for example, can make calls, negotiate bills, cancel subscriptions, and pursue refunds on a user’s behalf. The lender must still process each interaction through systems designed for a much higher cost of creation.
One side standardizes to control cost and risk; the other customizes every request. The answer is not to stop automating, but to design for machine-generated demand that can arrive continuously and at almost no cost to the sender.
3. Optimize: They learn to pass your gates
External agents do not only ask for more. They can shape the interaction to fit the signals your organization expects to see.
An appeal can be rewritten around the language most likely to trigger review. A mortgage application can be refined to clear screening criteria. Supporting documents can be polished until they look authoritative, whether or not the underlying claim deserves the same confidence.
This exposes a deeper problem: effort stopped being proof.
Writing a formal complaint, preparing a strong application, or filing an appeal once took time. That effort acted as a filter. The polish of the result served as a rough proxy for seriousness, sincerity, and stakes.
Now a flawless, maximum-pressure artifact can be produced in seconds. The hill is gone. Polish no longer signals what it used to, yet many organizational instincts for “serious versus noise” remain calibrated to that disappearing signal.
From Customer Experience to Agent Experience
For two decades, enterprises have invested in Customer Experience on the quiet assumption that a human sat on the other end of every interaction.
Increasingly, the customer may not experience your organization directly. Their agent does—and an intermediary AI layer may interpret the experience for them.
That makes Agent Experience, or AX, an emerging discipline: how external agents discover, interpret, navigate, and represent your organization. AX does not replace CX. It increasingly shapes it.
The question is no longer only whether your agent works. It is whether your data, policies, processes, and channels still work when another agent arrives.
What leaders should do now
The response should follow the same three-move framework:
Move | Leadership question | First action |
Judge | What do external agents say about us before direct engagement? | Test high-value prompts across major AI interfaces. Compare the answers with verified facts and identify missing context. |
Manufacture | Where can machine-generated demand trigger disproportionate human work? | Stress-test complaints, appeals, claims, retention, and support journeys. Measure the downstream work each request creates. |
Optimize | Which decisions still treat effort or polish as evidence of legitimacy? | Review screening signals. Add provenance, verification, context, and targeted human review where presentation can no longer be trusted. |
Across all three moves, AX requires consistency: policies, criteria, status information, and next steps must be current, clear, and usable across the channels external agents encounter. The goal is to make legitimate machine-mediated interactions easier to verify and resolve without allowing generated volume to overwhelm operations.
Every enterprise must still ask what its agents can do for customers. But that is only half the strategy.
Ask not what your agent can do for your customers. Ask what you can do for their agents.
Resources:
Gartner. (2025, February 10). Traditional customer service channels are losing ground to mobile and AI innovations. https://www.gartner.com/en/newsroom/press-releases/2025-02-10-traditional-customer-service-channels-are-losing-ground-to-mobile-and-ai-innovations