The Cost of Manual Quality Control Is Not a Line Item, Which Is Why It Never Gets Managed
Ask a mortgage operations leader what origination costs and you will get a number. Ask what quality control costs per loan and the answer is usually a headcount, a vendor invoice, or a pause.
This is not an oversight. It reflects something true about how the cost is structured: most of it does not appear where it is incurred.
What is measured
Fully loaded loan production expense: commissions, compensation, occupancy, equipment, corporate allocations and other production costs: reached $11,109 per loan in Q3 2025, up from $10,965 in Q2. [1] Costs rose by close to $800 per loan in Q1 2026 before declining again in Q2. [1] For depositories specifically, retail cost to originate averaged $16,320 per loan across 2025, a 4% improvement on the 2023 study high of $17,071. [1]
Against a long-run average of $7,799 per loan since Q1 2008, current costs run well above the historical norm. [1]
These are the figures that get quoted, and they are the wrong ones for this question. They describe total production cost. Quality control is inside them, undifferentiated, alongside everything else.
There is no widely published, methodologically consistent per-loan QC cost figure for the US mortgage industry. That absence is worth sitting with, because it is not a data gap so much as a structural feature: QC cost is distributed across departments in a way that resists being totalled.
The four places the cost actually sits
Direct review labour. The visible portion: reviewers, QC managers, vendor fees. This is the part that appears in a budget and the part that gets cut when volume falls. It is also, in most organisations, the smallest of the four.
Rework and cure. A finding generates work elsewhere: conditions re-cleared, documents re-requested, files reopened after closing. That work is performed by processing, underwriting and closing staff, and it is booked to those functions. From the QC budget's perspective it is invisible. From the company's perspective it is often the largest component.
Escaped defects. The defects that reach an investor arrive as repurchase demands, at an estimated average cost of $32,288 per loan. [2] This cost is booked to secondary marketing or to a reserve, usually months or years after the review that missed it, and by then the causal link to a QC decision is no longer traceable in the accounts.
Foregone coverage. The cost that never appears anywhere. When volume rises and the team does not, the sample shrinks or the turn time slips. Shrinking the sample has no accounting entry at all: it registers as the QC function absorbing volume efficiently.
Only the first of these lands in the QC budget. So the reported cost of quality control is roughly the one component that scales predictably, while the three that scale badly are attributed elsewhere.
Why this produces a specific, repeatable mistake
When a cost is measured in one place and incurred in four, the optimisation applied to the measured part looks like a success on its own terms.
Reduce review headcount, and direct QC cost falls immediately and visibly. Rework cost rises in other departments, where it is diluted among their other work. Escaped defects rise, but surface after a lag long enough that they attribute to market conditions or to a particular vintage rather than to a staffing decision. Coverage falls, and nothing records it.
Every one of those consequences is real. None of them shows up as an increase in the cost of quality control. The measurement boundary makes a bad trade look like a good one, reliably, and the feedback that would correct it arrives too late and too diffusely to be connected to the decision.
The scaling property that matters
Manual review scales linearly with headcount, and that is the property that determines what happens under stress.
When volume doubles, a linear-cost function offers three options: double the team, shrink the sample, or extend the turn time. Under actual conditions, volume moving faster than hiring, the sample shrinks. It is the only lever available immediately, and it is the one with no accounting consequence.
This is why the composition of defects keeps moving while the headline rate does not. The critical defect rate was 1.50% for 2025 against 1.52% the prior year, effectively flat. Inside it, Legal, Regulatory and Compliance rose roughly 30% in Q4 to 24.66%, its third consecutive quarterly increase; Borrower and Mortgage Eligibility rose 291.58% across the year and Credit rose 166.13%. [3]
A flat headline rate covering category movement of that magnitude is worth investigating. It may indicate that risk is relocating faster than the review programme is following it. It is not, on its own, evidence that sampling has deteriorated: a well-designed random sample estimates an overall rate perfectly well while the composition underneath it shifts. Treat the composition change as a question to investigate, not as a finding.
What a different cost structure changes
The argument for automating validation is usually made as labour substitution: fewer reviewers, lower direct cost. That framing concedes the wrong ground, because it accepts the same narrow measurement boundary that caused the problem.
The more consequential change is to the shape of the cost function. When an automated check does not scale with headcount, the depth of that check stops being the variable that silently absorbs volume growth.
This does not remove the QC sample, and any vendor telling you otherwise is describing a compliance problem. Fannie Mae requires both random and discretionary selection in a lender's post-closing quality control programme, and discretionary reviews supplement the random sample rather than replace it. [4] The random sample is what produces a defensible estimate of the defect rate; automation across the whole population does not substitute for it, because a population-wide automated check and a full-file human review are not the same evidence.
What changes is what the reviewer sees when the sampled file reaches them, and what is known about the files outside the sample. Automated cross-document checks can run over the whole population and surface candidates that a random draw would be unlikely to reach, which is the input a discretionary selection is supposed to be built from. The required random sample continues alongside it.
That does not eliminate the reviewer either. It removes the pressure to reduce depth, which is the decision that generates the three invisible costs.
The measurement worth building first
Before evaluating any system, including an automated one, it is worth constructing the number that does not currently exist: a fully loaded QC cost per loan that includes rework performed outside the QC function, an allocation of repurchase and indemnification cost against the vintages that produced it, and an explicit statement of coverage as a percentage of production.
Most organisations find that the third figure has moved without anyone deciding it should.
That number is hard to build and imprecise once built. It is still more useful than the one currently in the budget, because it measures the thing being managed rather than the thing that is easy to count.
Sources
- Mortgage Bankers Association, Quarterly Mortgage Bankers Performance Report Q3 2025, total loan production expenses per loan ($11,109); and MBA production-channel data for 2025, retail channel, depository institutions ($16,320). Both are total cost to produce one loan, not QC cost. Figures are quoted from MBA's published summaries; confirm against the current release before using them in a business case.
- STRATMOR Group, Unpacking the drivers and costs of GSE repurchase demands, reported by National Mortgage News. The $32,288 figure is a study estimate of cost per repurchase demand, not per completed repurchase; income and appraisal together account for 57% of demands in that study. A category share does not establish that those demands were preventable.
- ACES Quality Management, Mortgage QC Industry Trends Report, Q4 and CY 2025. Critical defect rate 1.50% for CY 2025 against 1.52% for CY 2024. Category figures are shares of critical defects, not defect incidence, and are drawn from the report's sample of reviewed loans.
- Fannie Mae Selling Guide, D1-3-01, Lender Post-Closing Quality Control Review Process. Requires both random and discretionary selection; discretionary reviews supplement the random sample rather than replace it.