Product Vision · 2026

The loan file is the interface.

Underwriting doesn't happen in a single form, tab, or conditions screen. It happens across an entire file — paystubs, W-2s, tax returns, leases, transcripts, appraisals, verification reports. If that's where the work lives, that's where the interface should live.

May 8, 2026 13 min read
Share

Most mortgage software still behaves as if underwriting were a field-entry problem. Open a tab. Update a value. Add a condition. Launch a calculator. Open a PDF in another window. Copy a number into the LOS. Ask a processor for a missing page. Repeat.

But that isn't how real underwriting works. Real underwriting happens across an entire file, where a fact in one document has to agree with a fact in another, and the hardest questions are almost never answered by a single screen. They're answered by evidence.

$1,700
per loan saved by lenders maximizing digital capabilities (Freddie Mac)1
~5
days cut from production timelines with digital tooling (Freddie Mac)1
4×
less likely to carry loan defects with digital tools (Freddie Mac)2
50%
of DU validation pilot lenders saw cost savings (Fannie Mae)3
$11,800
average production cost per loan, Q2 2025 (Freddie Mac)1
1000+
lenders already using GSE income-calculator tools (Freddie Mac)2

The form-filling fallacy

Agency policy makes the point plain. Fannie Mae's rules for bonus, commission, overtime, and tip income require lenders to compare a recent paystub with prior W-2s, analyze the income trend, and calculate qualifying income differently depending on whether the trend is stable, increasing, or declining.4 For self-employed borrowers, the core task is to determine the income that can actually be relied on — analyzing distributions, business viability, and stability of earnings, with a written analysis retained in the file.5 Underwriting here is not "read one document." It is "reason across time, documents, and context."

Yet most existing tools split that job into fragments. One system stores the record. Another validates income. Another runs quality checks. Another extracts data. A chatbot can summarize a PDF. What's usually missing is a single interface that lets an underwriter interrogate the whole file as a living network of evidence.8

"The old interface stores documents. The new interface understands the relationships between them."

— The thesis, in one line

The file is a network of evidence

Imagine opening a file and seeing not a stack of attachments, but an evidence graph. The paystub connects to the W-2. The W-2 connects to the transcript. The lease connects to Schedule E. Each document carries claims that must agree with claims elsewhere. Explore the file below — and switch on contradiction detection to see the pairs that disagree.

Synthetic loan file · evidence graph
Hover or focus a document to trace what it proves and what it must agree with
Synthetic example for illustration. Evidence relationships follow agency documentation patterns for wage, self-employed, and rental income.

Instead of asking "where is the right document?", the underwriter asks "what does the file prove?" That's a more useful question — and much closer to how experienced humans already think.4

Where the file disagrees with itself

The differentiated value isn't OCR or doc-chat — it's catching the moment two documents quietly disagree. Pick a scenario and watch claims get extracted, compared, and resolved into a single flagged conflict, with the policy context one click away.

1Document snippets
2Extracted claims
3Conflict detected
Synthetic snippets for illustration. Each scenario maps to a real agency reconciliation requirement.

The takeaway

A contradiction caught on intake costs one touch. The same contradiction caught in QC costs a condition, a re-upload, a recalculation — and sometimes a restart. The interface that reads documents against each other is the one that protects the margin.

Two tools for one job

Today's stack is optimized for workflow and records. ICE Encompass is sold as the system of record for the mortgage lifecycle; extraction vendors like Amazon Textract and Google Document AI are superb at OCR and classification.8,9,10 What none of them is built to do is reason across the whole file. Toggle the role below to see the difference between managing steps and interrogating evidence.

Illustrative. The legacy side optimizes for steps and records; the file-native side optimizes for cited answers.

Complexity compounds

The reason this matters more every year: complexity doesn't rise linearly. When income sources interact — a self-employed borrower who also owns rentals — the document count and the interpretation load climb together. The slowest steps aren't OCR problems; they're interpretation problems.

14
W-2 conforming
19
Variable income
25
Rental refi
33
Self-employed
41
Self-emp. + rental
49
Complex jumbo
Documents per file rise nonlinearly when income sources interact.
Legacy: first decision-ready pass
155min
File-native: first decision-ready pass
38min
Illustrative model for storytelling — document counts and times are modeled estimates calibrated to agency documentation complexity, not measured benchmarks.

It has to show its work

Speed is only half the story; trust is the other half. The CFPB has made clear that lenders cannot hide behind black-box complexity when they take adverse action — if a model influences a credit decision, the creditor still has to provide specific reasons.6 NIST's AI Risk Management Framework points the same way, centering explainability, transparency, accountability, and reliability.7 In mortgage, that means the interface can't just produce an answer. It has to show its work.

The market is already moving toward data over documents. Fannie Mae's DU Validation Service validates assets, income, and employment digitally while reducing collected documents and accelerating clear-to-close.3 Freddie Mac positions its Income Calculator and AIM tools as a spectrum that moves lenders toward instant results and earlier certainty.2 An evidence-preserving interface is how those gains become defensible at the level where underwriters actually work: the file.

Ask the file anything

This is where the idea stops being conceptual. Below is a read-only, synthetic loan file. Ask it a question and it returns a conclusion, the line-item evidence behind it, a confidence label, and a reminder that the human stays in control.

Sandbox file
Rivera, M. — Conforming purchase
Synthetic · read-only · 24 documents
1003PaystubW-2Sched ELeaseVOEBank4506-C
Pick a question to interrogate the file.
Synthetic, read-only demonstration. No live borrower data; answers are precomputed for illustration.

What earlier certainty is worth

Freddie Mac says lenders maximizing digital capabilities save roughly $1,700 per loan and shorten timelines by about five days.1 Those aren't Power Underwriter claims — they're proof that moving from document handling to evidence-driven work changes the economics. Adjust the inputs below to model what the document-review portion alone could be worth in your shop.

600
24
4.6
$58
10%
38
Hours saved / month
0
Labor saved / month
$0
Estimated annual savings$0
Underwriter capacity unlocked0 FTE
Implied savings per loan$0
Illustrative model. Assumes a 55% reduction in document-review time and a 45% reduction in rework — conservative defaults that keep per-loan savings well below Freddie Mac's published ~$1,700 figure. Replace with your own telemetry before relying on it.
Interactive estimate. Inputs are editable; outputs update live.

Frequently asked questions

It means the underwriting workspace should be the loan file itself — read as a connected network of evidence — rather than a maze of tabs, fields, calculators, and condition queues. Because agency rules require reconciling facts across paystubs, W-2s, tax returns, leases, Schedule E, and verification reports, the work is inherently cross-document. The interface should let an underwriter interrogate the whole file and get cited, traceable answers.
Extraction tools like Amazon Textract and Google Document AI are excellent at OCR, classification, and pulling fields out of documents — but they aren't built to reason across an entire mortgage file, apply agency policy, detect cross-document contradictions, or produce a defensible decision narrative. A loan-file-native interface adds the reasoning and evidence layer on top of extraction.
No — it's decision support. The system extracts claims, reconciles them across documents, surfaces contradictions, and preserves an evidence trail, but the human underwriter makes the decision. Explainability and adverse-action rules require a documented human judgment, so the interface is designed to show its work, not to decide on its own.
Every answer is tied to source pages with a reasoning trail. CFPB Circular 2022-03 makes clear that creditors using complex algorithms must still provide specific adverse-action reasons and cannot hide behind model complexity, and NIST's AI Risk Management Framework centers explainability, transparency, accountability, and reliability. An evidence-preserving interface is built to meet those expectations directly.

Sources

Figures attributed to Fannie Mae, Freddie Mac, the CFPB, and NIST are drawn from the publications below. The document counts, review times, and ROI projections in this article are illustrative modeled estimates for storytelling — calibrated to agency documentation complexity and reported GSE outcomes, not measured Power Underwriter benchmarks.

  1. Freddie Mac — 2025 Updates to the Cost to Originate StudyUp to ~$1,700 per-loan savings with LPA digital capabilities; ~$11,800 average production cost per loan, Q2 2025; shorter production timelines.
  2. Freddie Mac — How Freddie Mac Is Powering Efficiency and Cost SavingsLoans using Freddie Mac digital tools are on average 4× less likely to carry defects; Income Calculator / AIM adoption across 1,000+ sellers.
  3. Fannie Mae — Desktop Underwriter (DU) Validation ServiceDigital validation of assets, income, and employment with fewer collected documents; 50% of pilot lenders saw cost savings; enhanced employer-name matching.
  4. Fannie Mae — Selling Guide, B3-3 Income AssessmentCross-document requirements for variable income trend analysis, employment history, and multiple concurrent income sources.
  5. Fannie Mae — Selling Guide, Income Calculator & self-employment analysisSelf-employed income requires written cash-flow analysis of business and personal tax data, retained in the file.
  6. CFPB — Circular 2022-03: Adverse-action notification with complex algorithmsCreditors using complex algorithms must still provide specific adverse-action reasons and cannot excuse opaque decisioning.
  7. NIST — AI Risk Management FrameworkTrustworthy-AI characteristics: validity, reliability, explainability, accountability, transparency, privacy; plus the Generative AI Profile.
  8. ICE Mortgage Technology — EncompassPositioned as the end-to-end loan origination system and system of record for the mortgage lifecycle.
  9. Google Cloud — Document AIDocument extraction, classification, and OCR — representative of extraction-first tooling.
  10. Amazon Web Services — Amazon TextractScalable OCR and structured-data extraction from documents — extraction-first, not loan-level reasoning.
See the file become the interface

Stop opening more tabs. Start interrogating the file.

Watch a synthetic loan file go from scattered PDFs to cited underwriting answers — on top of the LOS your team already works in.

Power Underwriter Research

AI underwriting for mortgage brokers & lenders

We build the AI underwriting assistant that reads complete loan files, calculates income and assets, reviews conditions, and generates structured reports — without leaving your loan origination system. This piece draws on Fannie Mae, Freddie Mac, CFPB, and NIST guidance; the interactive figures use synthetic data and illustrative modeling, not measured benchmarks.