Skip to content
Sei

Where AI Actually Takes Cost Out of a Mortgage: The Touches Worth Removing, the $11,000 Loan, and the Controls You Cannot Automate Away

6 min readPranay Shetty
Share

The Number That Starts Every Conversation

Every mortgage production leader I talk to opens with the same number. The fully loaded cost to originate a loan has been punishing. The MBA's Quarterly Mortgage Bankers Performance Report put per-loan production cost at $12,593 in the first quarter of 2024, $10,806 in the second, and $11,230 in the fourth. Set that against a long-run average nearer $7,600 across the fifteen years the MBA has tracked it, and the size of the problem is clear. The cost is high, most of it is people, and the pressure to bring it down with automation is real.

The instinct that follows is to point AI at the process and expect the cost curve to bend. That instinct produces disappointing results, because it treats every human touch in a mortgage as cost to be removed, and some of those touches are not cost. They are controls. They exist because a regulator, an investor, or a hard-won lesson put them there, and removing them does not save money, it moves the cost to a place that shows up later and larger.

The useful framing is not "how much of the mortgage can AI do." It is "which touches are cost, which are controls, and which are controls that AI can make cheaper without weakening." Getting that sorting right is the difference between an AI program that lowers cost to originate and one that lowers it on paper while raising repurchase risk, examination risk, and the cost of the cleanup.

The Touches That Are Pure Cost

Some of the human time in a mortgage exists only because the tooling was bad. It is coordination, chasing, and rekeying that produces nothing a borrower or a regulator values, and it is the first place AI belongs.

The largest example is document chase. A processor spending an afternoon calling a borrower for the bank statement that was missing at intake, then calling again for the page that was cut off, is doing work that exists because the front door let an incomplete file through. Move that work to the point-of-sale intake, where an agent reads the borrower's profile and asks for exactly the documents it implies while the borrower is still in the session, and the chase mostly disappears. The cost was never the document, it was the round trips, and the round trips are avoidable.

Status communication is the second. A meaningful share of a loan officer's and processor's day goes to answering "where is my loan," from borrowers, from real estate agents, from the borrower's spouse. That is real time spent on a question the system already knows the answer to. An agent that answers status accurately, on the borrower's channel, at the moment the borrower asks, removes that time without removing anything anyone was trying to protect.

Internal reconciliation is the third. Time spent making the number on one system match the number on another, catching a rate that did not carry from the lock to the disclosure, confirming the figures on the application and the closing disclosure agree, is coordination cost. An agent that checks data integrity across systems on every file does it faster and more completely than a person doing it on a sample, and nothing of value is lost when the reconciliation stops being manual.

None of these touches is a control. No investor requires that a human personally chase the missing bank statement, no regulator is protected by a loan officer reading a status update aloud. These are cost, and cost is what AI should take.

The Touches That Are Controls Wearing the Costume of Cost

Then there are the touches that look like cost and are actually controls, and mistaking one for the other is how AI programs get lenders in trouble.

The underwriting decision is the clearest. It is expensive, it is slow, and it is tempting to automate it away entirely, and it is a control. Someone has to make the reasonable, good-faith ability-to-repay determination the rule assigns to the creditor, and that determination carries liability that runs for the life of the loan. AI can and should do the work that feeds the decision, recompute the income, assemble the file, surface the trend, so the underwriter decides rather than assembles. What it cannot do is be the decision, because the decision is the control and the accountability that attaches to it does not transfer to a model.

The adverse-action notice is a second one that hides as cost. Denying a loan and explaining why in the specific terms ECOA and the FCRA require is time-consuming, and it is a consumer-protection control. An AI agent can draft the notice from the actual reasons the file produced, which makes it faster and more accurate, but a human has to stand behind the reasons, because a wrong or vague adverse-action reason is a legal exposure, not a productivity metric. The control is that the denial is explained truthfully and specifically, and that control has to survive the automation.

Fair-lending review is the third. Testing whether similar borrowers are treated similarly is expensive, it produces no loan directly, and it is the control that keeps the whole operation out of the worst kind of enforcement. AI actually makes this control stronger, because a consistent, documented process is testable in a way that a pile of individual judgments is not. But the answer to cost pressure is never to do less fair-lending testing. It is to make the testing cheaper by making the underlying process consistent, which is a different move entirely.

The trap is that all three of these read as expensive human touches on a cost report, and a program measured only on cost per loan will point AI straight at them. The result is a lower number this quarter and a repurchase demand, a UDAAP finding, or a fair-lending exam two years later that costs a multiple of what was saved.

The Touches That Are Controls AI Makes Cheaper

The most valuable category is the one most programs miss: controls that AI does not remove but makes dramatically cheaper to run at full strength. This is where the cost-to-originate math actually improves without any control weakening, and it is where I steer every production leader who asks.

Quality control is the prime example. Pre-funding QC is a control that most lenders run as a small sample because a human team cannot review the whole pipeline before it funds. An agent can review every file before the wire, which does not weaken the control, it strengthens it, and it converts defects that used to be caught after delivery, when the only remedy is repurchase, into defects caught before closing, when the fix is cheap. The control gets stronger and the cost goes down at the same time, which is the outcome the whole exercise is supposed to produce.

Consistency in judgment-adjacent work is the other. Self-employed income calculation is slow and inconsistent when it lives in senior underwriters' spreadsheets. An agent that runs the same documented method on every file frees those underwriters to spend their time on the reasonableness call that is genuinely theirs, and it produces a consistent record that makes the fair-lending control testable. The judgment stays human, the arithmetic gets automated, and the control the arithmetic supports gets easier to prove.

The pattern across this category is the same. The control does not move. The human keeps the judgment and the accountability. What moves is the mechanical work around the control, and moving it is what lets the control run at a coverage and a consistency a human team could never afford. That is the cost-to-originate improvement worth chasing, because it survives the audit.

How I Tell the Categories Apart

The test I use with production teams is a single question asked of each touch: if a regulator or an investor asked why this step exists, what is the answer? If the answer is "it exists because our tooling made us do it by hand," the touch is cost and AI should take it. If the answer is "it exists because someone has to be accountable for this decision," the touch is a control and the human accountability stays, though AI can do the work that feeds it. If the answer is "it exists to protect the borrower or the investor and we run it thin because it is expensive," the touch is a control AI can make cheaper at full strength, and that is the best place to invest.

The question sounds simple and it sorts almost every touch correctly, because it forces the distinction between work that produces value and work that produces defense. Cost-cutting that removes defense is not cost-cutting, it is deferral, and the deferred bill in mortgage comes with interest in the form of repurchases and enforcement.

The Honest Read

Cost to originate above $11,000 a loan is a real problem, and AI is a real part of the answer. The programs that get it wrong treat every human touch as cost and point AI at the whole process, which lowers the number this quarter and raises the risk that shows up later. The programs that get it right sort the touches first: they let AI take the coordination and the chasing that were never controls, they keep the underwriting decision and the adverse-action notice and the fair-lending judgment with the humans accountable for them, and they invest hardest in the controls AI makes cheaper to run at full strength, quality control and consistent income analysis chief among them.

At Sei, we build for that sorting rather than against it, because a mortgage operation that lowers its cost by weakening its controls has not lowered its cost, it has hidden it. The number worth improving is the one that stays improved after the exam. Take the touches that are pure cost, assist the touches that are controls, and make the controls you keep cheaper to run at full coverage. That is where AI actually takes cost out of a mortgage, and it is the only version of the savings that lasts.

Pranay Shetty

Pranay Shetty

CEO & Co-Founder

Related Articles

Mortgage

When Rates Drop, Your Servicing Portfolio Gets Called by Someone: Refinance Recapture With an AI Voice Agent, and the Tests That Decide Who It Calls

A rate move puts thousands of loans in the money in the same week, and the servicer has the data to reach those borrowers first. Before an AI agent dials any of them, it has to know which loans can actually close: VA's seasoning and recoupment tests under 38 U.S.C. 3709, the early payoff window on loans you sold, which consent covers a marketing call, and what the agent may and may not say about savings. What we learned running a voice agent on a lender's funnel, applied to the recapture campaign.

6 min read
Read more
Mortgage

A Fraud Alert on the Credit Report Is a Legal Stop: What FCRA Section 605A Requires Before a Mortgage Proceeds, and Who Makes the Call

An initial, extended or active duty alert on a tri-merge report restricts the lender by statute, and the verification has to go to the phone number in the alert, which is not always the one on the application. Address discrepancy notices carry their own duty under Regulation V. A technical account of how our agents read alerts on every credit pull, what they compare, and why we decided the verification call is made by a person.

6 min read
Read more
Mortgage

Social Security, Pensions and the Gross-Up: Calculating Nontaxable and Retirement Income Without Counting It Twice

The same monthly benefit produces a different qualifying income under Fannie Mae, FHA and VA, and the gross-up is applied to the whole check far more often than the rules allow. A technical account of how our income agent reads the taxable fraction off the tax return, reconciles the award letter against the deposit, tests continuance from documented terms rather than age, and where Regulation B draws the line on what an underwriter may ask.

5 min read
Read more

You Ain't Seen Nothin' Yet

Book a Demo

Pack up some of your complex historical files — any loan type, any investor. We run them through intake, income and condition clearing, and in 30 minutes you see every condition we created and cleared efficiently for your own team, and why.

  • Any loan type, any agency guideline or custom investor overlays.
  • Every finding cited to the guideline or document it came from

Case study highlight

Loan review fell from 4 hours 50 minutes to 47 minutes per file, and 78% of items now clear at high confidence — so QC touches only the exceptions.

Pre-Close QC Agent · Better Mortgage— read the case study

Please provide your full name so we know how to address you.

Tell us which company you represent so we can personalise our response.

Use your work email. One of our experts will get back to you.

Which desks would you like to discuss?*

Choose the desks you’d like us to cover. Pick “Not sure yet” if you’d rather we worked it out on the call.

Roughly, so we bring the right person to the call.

Complete the verification to submit the form.