The Volume Is Repeatable, Which Is Exactly the Trap
Pull a week of calls into a servicing shop and the pattern is stark. The large majority are a handful of questions asked over and over: why did my payment change, what is going on with my escrow, can I get a payoff figure, I do not understand my statement, what are my options because I am behind. A servicing operation that staffs for that volume is staffing a lot of people to say the same five things, which is exactly the kind of work that looks ready to automate.
It is, and the reason most attempts at it go wrong is that each of those five questions has a trapdoor underneath it. The question is routine. The answer is not, because a servicing answer is a regulated statement about what a borrower owes, and a casual version of it either misstates the number or walks straight past a sentence the borrower said that was, legally, something the servicer now has to act on. The difficulty was never answering the question. It is handling the volume at the quality the rules demand and catching the moment a simple question stopped being simple.
Two of the Five Are Statements You Can Get Wrong
Take the payment question and the statement question together, because they are the ones where a wrong answer is a compliance event, not just a bad customer experience. A borrower asks why their payment went up. The honest answer usually involves an escrow analysis, a change in taxes or insurance, the way a shortage is being spread, and the new payment that results. Every one of those numbers is a specific figure tied to that loan, and an agent that answers from a generic script instead of from the actual analysis will eventually tell a borrower something that is not true about what they owe.
That is the line that matters, because an agent that confidently states the wrong payment amount or mischaracterizes why it changed is not giving bad service, it is making a statement that can mislead a borrower about their mortgage, which is the territory the CFPB polices as an unfair, deceptive, or abusive act or practice under its UDAAP authority. Regulation X reinforces the same point from the operations side: the servicing policies-and-procedures rule at 12 CFR 1024.38 requires a servicer to have procedures reasonably designed to provide accurate and timely information to borrowers. So the design rule for an agent answering these two questions is that it never answers from a generic script. It answers from the loan, reading the same authoritative figures the statement and the escrow analysis were generated from, or it does not answer and routes to a person. A plausible-sounding number is the one failure the agent is not allowed to produce, because on a mortgage a plausible wrong number is worse than no answer.
One of Them Starts a Clock the Moment the Words Change
The options question, from a borrower who is behind, is the one where the agent has to listen for a shift in what the borrower is actually saying, because the same conversation can cross from a general question into a request that carries a legal deadline. A borrower asking in the abstract what programs exist is a question. A borrower who says they lost their job and cannot make the payment has raised a hardship, and the servicer's loss mitigation obligations under Regulation X begin to attach. The agent that treats both as the same FAQ either floods a curious borrower with a formal process they did not ask for or, far worse, hears a genuine hardship and answers it as trivia while a clock the borrower does not know about fails to start.
The same trapdoor sits under the payment and escrow questions in a different form. A borrower who says "you charged me for insurance I already have" or "my payment is wrong and I want it fixed" has not asked a question, they have asserted an error, and that assertion can be a notice of error under 12 CFR 1024.35, which obligates the servicer to acknowledge within five days and resolve within a defined window. A request for specific information about the loan can be a request for information under 12 CFR 1024.36, with its own timeline. Neither of those arrives labeled. They arrive inside an ordinary-sounding call, and the single most important thing the agent does is recognize the sentence that converted a question into an obligation.
Recognition Is the Product, Not the Answer
So the design we build toward puts the weight on recognition rather than on answering, because the answering is the easy part and the recognizing is where the compliance lives. The agent is built to classify what the borrower is actually doing on every turn of the call, and the classification that matters is not the topic, it is whether the borrower has crossed from asking into asserting or requesting. The moment a borrower asserts an error, the agent stops treating the call as an FAQ and moves it onto the error-resolution track: it logs the assertion, acknowledges it in the form the rule requires, and starts the clock rather than answering casually and letting the timeline run unacknowledged. The moment a borrower raises a hardship, the agent captures it and routes into the loss mitigation path instead of reciting program names.
This is the same discipline the servicing rules were written to force, built into the agent as behavior. A borrower never has to know the words "notice of error" for the protection to apply, which means the servicer, or the agent standing in for it, has to catch the substance regardless of the words. An agent that answers the routine version well but misses the moment the call became a notice of error is not a good servicing agent that occasionally slips. It is a compliance liability that happens to have good uptime, because the obligation it missed had a deadline that is now running silently.
Where the Agent Stops
The boundary is drawn so that the agent handles the repeatable informational volume and hands off anything that becomes a decision or a dispute. It answers the payment, escrow, payoff, and statement questions from the authoritative figures, which covers the bulk of the calls and covers them accurately because the numbers come from the loan rather than a script. It recognizes and routes the error assertions and the hardships, which are the calls that carry a clock. And it stops at anything that requires the servicer to exercise judgment or make a determination, a workout decision, a dispute resolution, an account adjustment, handing those to a person with the conversation and the context attached rather than improvising a resolution it has no authority to give.
The honest version of this boundary is that it keeps the agent on the side of the line where being helpful and being compliant are the same thing. Answering what a borrower owes, accurately, from the loan is both good service and a satisfied accuracy obligation. Recognizing a notice of error and starting the clock is both responsive and a met deadline. The agent is valuable precisely where those two goals coincide, and it is built to stop before the point where they could diverge.
What I Tell Heads of Servicing
The pitch for automating the servicing line is usually about deflection and cost per call, and that framing is how it gets built wrong, because a shop optimizing deflection will happily deflect a notice of error into a satisfied-looking FAQ answer and discover the missed deadline at the next exam. The right frame is that the agent exists to handle the repeatable volume at a quality a human under time pressure often cannot sustain, every payment question answered from the actual analysis rather than a guess, every error assertion caught and clocked the same way every time, which is more consistent than a call center having a busy Monday.
At Sei we build the servicing agent so that accuracy and recognition come first and deflection is a result rather than a goal, because the five questions that make up most of the volume are automatable and the trapdoors under them are exactly the places a deflection-first design falls through. Handle the routine calls from the loan, catch the sentence that turns a question into an obligation, and stop before the decisions that belong to a person, and you get the volume relief without the finding that usually comes attached to it.
Pranay Shetty
CEO & Co-Founder