A mortgage can move through a POS, LOS, AUS, pricing engine, document platform and several workflow tools and still consume hours of human attention.
That should bother us.
The industry has spent years improving individual steps in the mortgage process. Yet the economics remain difficult. According to the <a href="https://www.mba.org/news-and-research/newsroom/news/2026/08/18/imbs-production-profits-increase-in-second-quarter-of-2026″ rel=”nofollow noopener” target=”_blank”>Mortgage Bankers Association, total loan production expenses averaged $10,936 per loan in the second quarter of 2026. That was an improvement from the first quarter, but still well above the historical average of $7,945.
We measure cost per loan, cycle time, pull-through, defects and loans per employee. All of those matter. But as AI begins to move deeper into mortgage operations, I think we are missing one important measure:
How much human attention does it take to complete a loan? Not headcount. Not touches for the sake of counting touches. Actual human attention.
Where the work really goes
A processor may spend a few minutes finding the right document, another few reconciling two values, then time checking whether a condition was satisfied. An underwriter may reopen documents because the information presented to them is incomplete or not trusted. A QC reviewer may repeat checks already performed upstream because the evidence behind the earlier decision is difficult to reconstruct.
None of these activities looks particularly expensive in isolation. Across thousands of loans, they become the operating model.
This is why simply adding AI to existing workflows may not materially change the economics. If AI extracts information but a person still validates every field, we may have improved extraction without reducing work. If a model produces 30 findings and a reviewer must investigate 25 of them to discover that only three matter, we have created another queue.
The better question is not, “How much can AI do?” It is, “What work still deserves human attention after AI has done what it can reliably do?”
Separate work from judgment
Most mortgage operational work can be thought of in three buckets.
The first is repeatable work: classifying documents, identifying missing items, extracting information, comparing values, checking dates and completing deterministic validations. Increasingly, technology should perform this work without requiring a person to confirm every successful step.
The second is exception work.Something is missing, inconsistent, ambiguous or outside a defined tolerance. This is where technology should narrow the problem before handing it to a person. A reviewer should see the issue, the relevant evidence and why it was surfaced, rather than starting again with the entire file.
The third is judgment.This is where experienced people create real value: interpreting policy, evaluating materiality, understanding context, resolving ambiguity and taking accountability for consequential decisions.
The objective of AI should not be to eliminate the third bucket. It should be to move the first bucket away from people and make the second bucket smaller and more precise.
Measure attention, not activity
This creates a different way to evaluate mortgage technology.
For each workflow, lenders could establish a baseline for the human minutes required per completed loan and then break those minutes into reasons: document handling, data reconciliation, routine validation, exception resolution, judgment and rework.
Then measure again after automation.
A useful implementation should reduce human-attention minutes without increasing defects, rework or downstream risk. It should also increase the percentage of routine checks completed without human intervention and reduce the number of low-value exceptions sent to reviewers.
This is more revealing than an accuracy number alone. A system can report 98% accuracy and still create limited operating leverage if reviewers must check every output. Conversely, a narrowly scoped system that reliably completes a defined set of checks without human involvement may produce a much larger economic benefit.
This metric also changes capacity planning. Instead of asking how many loans one employee can process, an operations leader can ask how much expert attention a given volume will require and where that attention is likely to be needed.
That becomes particularly important as volumes change. Technology should allow capacity to flex without forcing expertise to scale linearly with every additional loan.
The next phase of mortgage automation
The mortgage industry does not need another race to remove people from the process. It needs a better way to use them.
Experienced processors, underwriters and QC professionals should not spend their time moving information between systems, repeating checks or searching through hundreds of pages to find the evidence behind an exception.
Their attention is expensive because it is valuable.
As AI becomes more capable, lenders should judge it by a simple operational test: After the technology has done its work, how much human attention is still required to complete the loan correctly?
If that number does not materially change, neither will the economics.
Alok Bansal is Founder and CEO of Logikality.
This column does not necessarily reflect the opinion of HousingWire’s editorial department and its owners. To contact the editor responsible for this piece: [email protected].
