The Underwriter's Inbox: Counting a Day at the Desk

Nobody records how many times an underwriter opens the same file. The credit decision gets recorded and the funding date gets recorded. The openings in between, the chase emails and the afternoon spent waiting on a bank statement appear nowhere, which makes them free in the reporting and expensive in the day.

Underwriting workflow efficiency is limited less by decision speed than by touch count. A file reopened nine times carries eight context reloads and eight waits between them. The composite day below is a model built from published interruption research, queueing theory and lending cost data, with every assumed figure marked as an assumption rather than presented as measurement.

Say plainly what this is. No lender, trade body or regulator publishes a step-level account of how an underwriting day is spent, and we went looking. What exists is research on interruption and task resumption, telemetry on message volume, queueing mathematics, and lending cost figures. The day below is assembled from those parts. Published figures are named in the sentence. Our own estimates are marked as assumptions, and there are more of those.

What does a day at an underwriting desk actually look like?

An underwriting desk is fragmented by default. Gloria Mark's 2005 CHI study of 24 information workers found an average of 11 minutes 4 seconds on a working sphere before switching, and 25 minutes 26 seconds before an interrupted one was resumed, among those resumed the same day. An assumed 450 minute day holds about 41 such stretches.

The fragmentation is not the underwriter's doing. Mark's 2005 study, run across more than 700 hours of observation, found 57.1% of work segments were interrupted, and that a worker visited 2.26 other pieces of work on average before returning. An underwriting desk has more legitimate reasons to interrupt than most: a funder calling about a file closing today, a broker who wants a verbal before he loses his client, a condition arriving out of order.

Message volume sets the floor under all of it. Microsoft's Work Trend Index, published 17 June 2025 from Microsoft 365 telemetry and a survey of 31,000 knowledge workers, reported the typical employee receiving 117 emails and 153 chat messages a weekday and being interrupted roughly every two minutes during core hours. Microsoft sells the software that generates those messages.

One interruption is cheap. Jackson, Dawson and Wilson put average recovery after an email interrupt at 64 seconds in 2001, from video of 15 employees at a single UK firm. Frequency is what makes it expensive.

Table 1: the composite underwriting day and the composite conditioned file. Rows marked Sourced carry a named published figure. Rows marked Assumption are our estimates, chosen to structure the argument, and are not measured lending data. Rows marked Derived are arithmetic on the rows above them. No published source breaks an underwriting day into touches per file, which is why the assumption rows outnumber the sourced ones.

#	Element	Model value	Basis
1	Length of the working day	450 minutes	Assumption
2	Average uninterrupted stretch on one item of work	11 min 4 s	Sourced: Mark, González and Harris, CHI 2005
3	Stretches available in the day	about 41	Derived from rows 1 and 2
4	Share of work stretches that were interrupted	57.1%	Sourced: Mark, González and Harris, CHI 2005
5	Stretches spent on credit files	26	Assumption
6	Stretches spent on messages, calls and meetings	15	Assumption
7	Distinct files touched in the day	14	Assumption
8	Messages arriving per person per weekday	117 email, 153 chat	Sourced (vendor research): Microsoft Work Trend Index, 17 June 2025
9	Recovery time after a single email interrupt	64 seconds	Sourced: Jackson, Dawson and Wilson, 2001
10	Elapsed time before an interrupted item is resumed, where it was resumed the same day	25 min 26 s	Sourced: Mark, González and Harris, CHI 2005
11	Other items visited before resuming	2.26	Sourced: Mark, González and Harris, CHI 2005
12	Touches in the life of one conditioned file	9	Assumption
13	Of those, touches that add new information	4	Assumption
14	Of those, touches that chase, confirm or re-read	5	Assumption
15	Reload time on re-entering a file	3 minutes	Assumption; direction sourced from Monk, Trafton and Boehm-Davis, 2008
16	Underwriter time per file across its life	about 100 minutes	Derived from rows 2 and 12
17	Share of that time spent reloading context	about 24%	Derived from rows 15 and 16
18	Wait for an applicant reply	1 business day	Assumption
19	Wait for a broker or dealer reply	3 hours	Assumption
19a	Split of the five chase touches	3 wait on an applicant, 2 on a broker	Assumption
20	Desk utilisation	85%	Assumption; the threshold itself sourced from Bagust, Place and Posnett, BMJ 1999
21	Queue multiplier at 85% utilisation	5.7	Derived from row 20 via Kingman's formula, 1961, with the variability term set to 1
22	Elapsed life of the file	about 5 business days	Derived from rows 12, 18, 19, 19a and 21
23	Share of elapsed life spent working on the file	under 5%	Derived from rows 16 and 22

How many touches does one file take?

Touches per file is the count of separate occasions a file is opened, worked and set down again before a decision is reached. No lender publishes a distribution of it. The composite here assumes nine touches on a file that needs two rounds of conditions, of which four add new information and five chase, confirm or re-read.

Nine is the load-bearing assumption here, so it deserves the scrutiny. It comes from a shape rather than a study: first read, a request for missing items, the partial response, a second request, the response to that, a verification callback, the write-up, a second signature, the funding package. A self-employed applicant with seasonal income can run well past nine.

Nine touches at Mark's measured 11 minutes 4 seconds gives roughly 100 minutes of underwriter time across a file's life. Around 24 of those minutes go to reloading a file already read, at an assumed three minutes per re-entry. A quarter of the work is re-reading the work.

The re-read is invisible to everyone except the person doing it, which is why it never gets designed out. Four information-adding touches is close to the floor for a properly conditioned credit file. Five chase touches is not a floor. It is a consequence of how the file was collected.

Where does the waiting actually happen?

Wait states dominate elapsed time in lending files because a busy desk queues nonlinearly. Kingman's 1961 approximation, published in Mathematical Proceedings of the Cambridge Philosophical Society, makes mean waiting time proportional to utilisation over one minus utilisation. At 85 percent load that term is 5.7. At 70 percent it is 2.3.

Read those two numbers together and a busy underwriting desk becomes arithmetic rather than temperament. A 15 point reduction in load cuts the waiting term by about 59 percent. The relationship runs the other way too, which is why a desk that feels fine in March feels impossible in May on a volume increase that looked modest on the board deck.

Health services research has argued about this threshold for twenty-five years. Bagust, Place and Posnett, in a stochastic simulation published in the BMJ in July 1999, found risks of bed shortages became discernible above roughly 85% average occupancy. Nathan Proudlove, in Health Services Management Research in 2020, argued that treating 85% as a universal target gets the logic backwards, since utilisation is an output of the access level a system wants. Neither was measured on a credit desk.

The other half of the wait sits outside the building. The composite assumes three of the five chase touches wait on an applicant and two wait on a broker, and one business day is generous for someone who is at work when the request arrives. Nothing in those hours is underwriting.

What does it cost to pick a file back up?

Picking a file back up costs time that no lending system records. Monk, Trafton and Boehm-Davis, writing in the Journal of Experimental Psychology: Applied in 2008, measured resumption lag at 1,548 milliseconds against a 949 millisecond uninterrupted baseline, and found it rising with both the duration of the interruption and its mental demand.

Their third experiment matters most here. Across 36 participants, resumption lag ran 1,322 milliseconds where the interruption imposed no secondary task, 1,605 where it imposed a tracking task, and 1,789 after a demanding one. Underwriting interruptions are rarely the empty kind.

Sophie Leroy's 2009 paper in Organizational Behavior and Human Decision Processes, volume 109, described the mechanism as attention residue: people carry an unfinished task into the next one and perform worse on it until they have genuinely put the first down. An underwriter holding four half-conditioned files is holding four residues.

Gloria Mark, Daniela Gudith and Ulrich Klocke found in a 2008 CHI experiment with 48 participants that interrupted work was completed faster than uninterrupted work, at 20.3 to 20.6 minutes against a 22.8 minute baseline, while measured stress rose from 6.92 to about 9.3 on a 20 point scale. People compensate, and the compensation comes out of the person rather than the queue.

None of this was conducted on credit files. The transfer is an analogy about mechanism, and the three minute reload assumption is the weakest number in Table 1. The underlying research measures a disruption cost of roughly 600 milliseconds on a laboratory task, so three minutes for a credit file is an extrapolation of two orders of magnitude.

What would one-touch files change about underwriting workflow efficiency?

One-touch files change underwriting workflow efficiency by removing work rather than by speeding anyone up. Take out the five chase touches and the waits attached to them and a nine-touch file becomes a two-touch file. On the composite figures in Table 1 that removes close to 80 minutes of desk time per file and most of the elapsed calendar.

Two touches at 11 minutes 4 seconds is about 22 minutes of work with one reload inside it. External waits go to zero because nothing is being chased. Queue time falls twice over, once because there are fewer touches and once because the desk is less loaded.

A desk that frees capacity takes more files, and utilisation drifts back toward wherever the staffing model puts it. Modelling a proportional fall in load from 0.85 to 0.19 would be arithmetic that never survives contact with a lending operation.

The money is well documented even where the workflow is not. The Mortgage Bankers Association's Quarterly Mortgage Bankers Performance Report for the first quarter of 2026, published 15 May 2026 across 324 companies, put total loan production expenses at $11,898 per loan against net production income of $727. At that margin, 80 minutes of avoidable desk time is not a rounding error.

Documentation is where the touches concentrate. Fannie Mae's Mortgage Lender Sentiment Survey, fielded to 170 senior mortgage executives in autumn 2024, found 71% describing variable income as difficult to use for qualification. The Federal Reserve Banks' 2026 Report on Employer Firms, published 3 March 2026 from 6,525 responses, found 36% of financing applicants received some or most of what they asked for rather than all of it. Both describe files opened more than twice.

What we could not verify

The core number in this article is unsourced. No published dataset gives touches per file for underwriting, in any product or any country. Cost per loan is published. Approval rates are published. How many times a human being opens a credit file is not, anywhere we could find.

We searched the Mortgage Bankers Association, Fannie Mae, the Federal Reserve Banks' small business work, ICE Mortgage Technology and Canadian federal sources. Cycle time is barely better served. ICE Mortgage Technology's Origination Insight Report put average days to close at 52 in March 2021, in a press release dated 27 April 2021, and that is the most recent openly published measurement of its kind we found.

Canadian workflow data is thinner again. Innovation, Science and Economic Development Canada's Small Business Credit Condition Trends, published 14 July 2026, reports 39% of Canadian small businesses requested external financing in 2025 and 97% of debt requests were approved, with no time-to-decision measure. CMHC's 2025 Mortgage Consumer Survey, fielded to 3,968 consumers in January 2025, records 2.7 lenders and 2.3 brokers contacted on average and does not measure documentation rounds. Every interruption and queueing figure here is American or British.

One more limit. Kingman's formula includes a variability term for arrivals and service times which we set to 1. Real underwriting queues are lumpier than that, month-end especially, so 5.7 more likely understates the waiting than overstates it.

Common questions

How many times does an underwriter open the same loan file?
No published dataset answers this. The composite model in this article assumes nine touches for a file requiring two rounds of conditions, four of which add new information. Nine is an assumption chosen to structure the argument and is explicitly not measured lending data.

How long does it take to get back into a file after an interruption?
Two published measurements bound it. Jackson, Dawson and Wilson found in 2001 that recovery after an email interrupt averaged 64 seconds, and Monk, Trafton and Boehm-Davis measured resumption lag at 1,548 milliseconds in 2008. Neither was measured on credit files.

Why does a busy underwriting desk wait so much longer than a quiet one?
Because waiting grows nonlinearly with load. Kingman's 1961 approximation makes mean waiting time proportional to utilisation divided by one minus utilisation, so the term is 2.3 at 70 percent load and 5.7 at 85 percent. Small volume increases produce large waiting increases.

What does a mortgage loan actually cost to produce?
The Mortgage Bankers Association's Quarterly Mortgage Bankers Performance Report for the first quarter of 2026, published 15 May 2026 and covering 324 companies, put total loan production expenses at $11,898 per loan and net production income at $727 per loan. Expenses rose close to $800 per loan over the prior quarter.

Is there Canadian data on underwriting workflow efficiency?
Not on workflow. Innovation, Science and Economic Development Canada publishes small business credit request and approval rates, most recently on 14 July 2026, and CMHC's 2025 Mortgage Consumer Survey covers consumer behaviour. Neither measures touches, documentation rounds or time to decision.

---

Related reading: per-step time allowances for intake, document collection over email, speed and diligence.

Carousel builds the intake layer that gets a file to the desk complete the first time. See how verification fits your flow