# A Mortgage Takes Forty-Odd Days and Spends Most of Them Waiting: Where the Time Actually Goes, and the Days an AI Agent Gives Back

*October 3, 2026 · 6 min read · Pranay Shetty*

> The industry still measures time to close in the mid-forties of days, and very little of that is work. It is a loan sitting in a queue between two people who are each busy with a different file. A production leader's map of where the dead time hides, which regulatory clocks keep running while the loan waits, and which gaps an AI agent closes without touching a decision anyone is accountable for.

## The Loan Is Not Slow Because the Work Is Hard

Every production leader I talk to can tell me their average time to close, and the number still lands in the mid-forties of days for most of them, in line with the time-to-close figures [ICE Mortgage Technology](https://www.icemortgagetechnology.com/) reports across the industry. What almost nobody can tell me is how many of those days were spent doing work, because the honest answer is embarrassing. A mortgage is not slow because underwriting is hard or because the documents take days to read. It is slow because the file spends most of its life waiting in a queue between two people who are each busy with a different file.

That distinction is the whole game for anyone trying to take time out of a loan. The work in a mortgage, reading the documents, calculating the income, clearing the condition, checking the disclosure, adds up to a few hours spread across the file. The forty-odd days are the gaps between those hours. If you point an automation program at the work, you shave minutes off a few-hour total and wonder why the cycle time did not move. If you point it at the gaps, the number moves, because the gaps are where the time actually is.

## The Three Places the Days Hide

Walk a slow loan backward from the closing table and the lost time clusters in the same three places almost every time.

The first is the document chase. A file stalls because something is missing, a bank statement, a page that was cut off, the explanation letter for a deposit, and the request to the borrower goes out, and then the loan sits. It sits overnight because the request went out at 4pm. It sits over the weekend because the borrower travels for work. It sits for three days because the processor who sent the request is working forty other files and does not circle back until the borrower has already gone quiet. None of that waiting is work. It is a loan parked against a human's attention that is spread too thin to return to it quickly.

The second is the handoff. A loan moves from the loan officer to processing, from processing to underwriting, from underwriting back to processing to clear conditions, and each of those boundaries is a queue. The file lands in someone's stack and waits its turn. The actual act of picking it up and doing the next step might take twenty minutes, but it happens eight hours after the file arrived because seven other files were ahead of it. Multiply that across the five or six handoffs in a normal loan and the queue time alone can be a week.

The third is the dead time around a decision that is waiting on an input. [Conditions clearing](/blog/ai-loan-processing-conditions-clearing-los-reg-b-incompleteness) is the worst offender, because a condition cannot clear until the evidence lands, and once it lands it still waits for someone to notice it landed, match it to the condition, and clear it. The evidence sat in a folder for two days before anyone connected it to the thing it satisfied. That is not review time. It is recognition time, the lag between an input arriving and a human realizing it arrived, and it is pure loss.

## The Clocks That Keep Running While the Loan Waits

The reason this dead time is worse than a cost-per-loan problem is that the regulatory clocks do not pause while the file sits in a queue. They run against the calendar, and the calendar does not care that the loan was waiting on a human's attention rather than being worked.

Regulation B gives a creditor thirty days from a completed application to notify the applicant of the action taken, under [12 CFR 1002.9](https://www.ecfr.gov/current/title-12/chapter-X/part-1002/section-1002.9). TRID runs its own clocks in parallel: the Loan Estimate has to reach the borrower within three business days of application, and the Closing Disclosure has to be received at least three business days before consummation, both under [12 CFR 1026.19](https://www.ecfr.gov/current/title-12/chapter-X/part-1026/subpart-C/section-1026.19). On the servicing side, Regulation X puts its own hard deadlines on error resolution and loss mitigation. Every one of those clocks keeps counting while a loan sits in a queue, which means dead time does not just slow the loan. It eats the margin a lender has to meet a deadline, and a file that burned two weeks in queues is a file with two fewer weeks of cushion before a disclosure clock turns a delay into a violation.

So the cycle-time problem and the compliance problem are the same problem seen from two angles. The days lost to queues are days the lender cannot spend, and a disclosure deadline that slips because the file sat too long is not a productivity miss, it is a finding.

## What an Agent Closes, and Why It Is the Gaps and Not the Work

An agent that holds a goal and works it continuously attacks the dead time directly, because the thing it does that a person cannot is not read faster. It is never go home. A loan does not wait overnight for the document request to go out, because the agent sends it the moment the gap appears. The evidence that lands at 11pm on a Saturday is recognized at 11pm on a Saturday, matched to the condition it satisfies, and the condition cleared, so the recognition lag that cost two days is gone. The handoff queue shrinks because the agent is not working forty other files in series, it is holding every file at once and moving each the instant it can move.

This is the part worth being precise about, because it is where automation programs either help or disappoint. The agent is not making the loan officer read faster or the underwriter decide quicker. The [ability-to-repay determination](/blog/reg-z-1026-43-ability-to-repay-qualified-mortgage-ai-underwriting) still belongs to the underwriter and still takes the judgment it takes. What the agent removes is the waiting around that decision, the time the file spent in a queue before the underwriter saw it and the time it spent after the input arrived before anyone noticed. Those gaps were never anyone's skilled work. They were the cost of human attention being a scarce, serial resource, and an agent that runs continuously and in parallel is the one thing that actually relieves it.

## The Measurement That Tells You If It Worked

If you are going to take time out of a loan, measure the right thing, because average time to close is too coarse to steer by. The measurement that tells the truth is touch-to-touch time, the gap between one action on the file and the next. Map a sample of your slow loans that way and the dead time shows itself: you will see a document request at 4pm Tuesday and the borrower's response at 9am Wednesday, which is fine, and then the next action on the file at 2pm Friday, which is two and a half days of a loan sitting on evidence that already arrived. That Friday gap is the recoverable time, and it is invisible in an average-days-to-close number that only sees the endpoints.

We instrument deployments this way on purpose, because a lender who measures only the endpoints cannot tell whether a faster close came from closing the gaps or from cutting a corner, and those two have very different consequences at the next exam. The honest version of a cycle-time improvement is one where the touch-to-touch gaps shrank and every control that was running still ran, each disclosure still cut on time, each decision still made by the person accountable for it. A faster number that came from skipping a step is not a cycle-time win. It is a deferred problem with a better dashboard.

## What I Tell Production Leaders

The instinct when cycle time is the problem is to look for the slow step and speed it up. Almost always there is no slow step. There is a fast step that happened late, because the file waited in a queue to reach it and waited again after its input arrived for someone to notice. The time is in the waiting, not the working, and the waiting is a function of human attention being scarce and serial. That is the constraint an agent that runs overnight and in parallel actually lifts, and it is why the cycle-time gains are real while the cost-cutting gains people chase by trimming the work are not.

At Sei we build for the gaps rather than the steps, because the forty-odd days in a mortgage were never forty days of work and the lender who speeds up the work is optimizing the wrong thing. Close the recognition lag, drain the handoff queues, and keep every control and every clock exactly where the rule puts it, and the loan closes faster for the only reason that survives an examination: it stopped waiting, and nothing it was supposed to do got skipped to make that happen.

---

_Source: [https://seiright.com/blog/mortgage-cycle-time-where-the-days-go-ai-agents-turn-time](https://seiright.com/blog/mortgage-cycle-time-where-the-days-go-ai-agents-turn-time) · Sei AI_