Every loan officer has met this borrower.
He is a Fort Campbell soldier, four years in, and he works overtime. Not occasional overtime. Real, every-week overtime, the kind that decides whether his family is looking at a $260,000 house or a $330,000 one.
So you build the file. His paystub year-to-date says one number. The automated verification of employment comes back with something a little different. Last year’s W-2 says a third thing. And his bank deposits do not quite match any of the three.
Nobody lied. Every one of those documents is accurate. They simply do not agree.
Gerald Green put a name to this problem in a piece in HousingWire this week, and it is the cleanest description of it I have read in 26 years of originating loans. One line stopped me: “We used to struggle to get the data. Now we increasingly struggle to determine what the data means when legitimate sources do not agree.”
That is the whole job now. Right there in one sentence.
The hard part moved, and a lot of us did not notice
For most of my career, the hard part was collection. You chased documents. You faxed. You called employers who never called back. Getting the paper was the work.
That problem is mostly solved. Payroll feeds, asset verification, tax transcripts, automated VOEs. A file that used to take three weeks to assemble now builds itself in an afternoon.
But solving collection did not solve understanding. It just moved the bottleneck. Now you have five authoritative sources and one decision to make.
Green draws a distinction here that I would write on the wall above your desk: a source can be authoritative for what it records without being sufficient for the fact you actually need to establish.
The paystub is authoritative. It tells you exactly what that employer paid on that date. It is not, by itself, enough to tell you what that borrower will reliably earn over the next twelve months. Those are two different questions. We have been treating them like one question for years.
Why this matters more this year than last year
Because something new is reading your file.
AI reads all of it. Faster than any human underwriter, and more completely. It compares records, catches the differences, and summarizes what probably happened. That is genuinely useful, and I am not here to talk anybody out of it.
But Green flags the failure mode, and this is the sentence that belongs on the wall of every underwriting shop in the country. AI should never be allowed to “turn conflicting evidence into an unquestioned fact simply because the workflow requires an answer.”
Read that last part again. Because the workflow requires an answer.
That is the risk. Not a system that is wrong. A system that is confident. The screen needs a qualifying income figure before it will advance, so it produces one. It averages. It smooths. It picks. And from that moment forward, that number travels through the entire file as settled fact, because nothing downstream has any idea it was ever in question.
Here is the uncomfortable part. Humans have been doing this for decades. Somebody picks the conservative number to be safe, and a family loses a house they could genuinely afford. Somebody picks the generous number because the deal needs to work, and that family closes on a payment that breaks them in month seven. Those look like opposite mistakes. They are the same mistake.
The best idea in the article: unresolved is an answer
Green writes something that cuts against every instinct in a commission business. Sometimes the correct result is unresolved.
Failure to prove something is not the same as proving the opposite.
I have built a good part of my career on that sentence without knowing how to say it out loud. It is why I take the files other people declined. A file that has not been proven yet is not a file that has been disproven. Those are wildly different things, and the difference is usually one document, one letter of explanation, or one phone call to the right person.
When a system flattens “we do not have enough yet” into “no,” real families pay for it. When it flattens the same thing into “yes,” they pay later, and worse.
What this means for your Monday morning
Green lists what trustworthy evidence architecture requires: provenance, temporal context, governed rules, reason codes, reconstructability. That is systems language. Here is the originator translation, and every one of these is free.
- Date everything, and keep periods separate. A March paystub and an August VOE are not describing the same reality. Stop stacking them as though they are.
- Say where the number came from. Not just “qualifying income $6,420.” Write the source. Write the method. Two lines of narrative in the file saves an hour of conditions later.
- Write down the why, not just the what. Why you used the 24-month average instead of 12. Why you excluded the bonus. Your reasoning is evidence too.
- Flag the conflict yourself, before the machine finds it. When your sources disagree, say so in the file and explain how you resolved it. An underwriter who sees that you already caught it reads the rest of your file differently.
- Stop averaging in your head. If you cannot show your work, you do not have a number. You have a guess wearing a number’s clothes.
Realtors, this is your problem too
You live the same thing with value.
The tax card says 1,850 square feet. The appraiser measures 1,790. Zillow says something else entirely. The listing says “approximately.” Your CMA pulled from all of it. Nobody lied. Those sources simply do not agree, and the gap can cost your client real money or cost you a contract.
The agent who says “here is the number, and here is exactly where it came from” is doing evidence adjudication, whether or not anybody calls it that. That is not extra work. That is the work.
Why I read this article as good news
Green’s closing point is the one I would build a business on. The advantage does not go to the company with the smartest AI. It goes to the system that can prove the facts the AI is acting on.
The proving is still human. The judgment about whether four documents actually support a conclusion, and the honesty to say “not yet” when they do not, is not a feature anybody is shipping next quarter.
Technology should make us more human, not less. And trust is still the greatest competitive advantage.
Here is your action step, and it will take you fifteen minutes. Pull one file out of your pipeline today. Find the qualifying income number. Then ask yourself whether you could hand that file to a stranger and have them reconstruct exactly how you got there, from the documents alone. If the answer is no, you just found the weakest link in your file. Go fix that one. Then do it on the next file, and the one after that, until it is simply how you work.
Knowledge is power. But proof is what closes loans.
Frequently Asked Questions
1. What does “evidence adjudication” actually mean in plain English?
It is the step where somebody decides whether the documents you have are good enough to rely on a specific fact. Not whether the borrower is approved. Just whether the evidence genuinely supports the number. Green’s own definition is careful about that line: it determines whether evidence sufficiently supports downstream reliance on a fact, “without deciding the borrower’s eligibility or credit outcome.”
2. Is this article saying we should not use AI in mortgage lending?
No, and neither am I. AI reading a full file is a real gain. The caution is narrower and more important: do not let a system convert a genuine disagreement between documents into a settled fact just because the next screen needs something in the box.
3. My borrower’s income documents do not match. Which one does the underwriter use?
It depends on what fact you are trying to establish, and that is the point. Agency guidelines give you methods for averaging and for treating variable income. What they cannot give you is a shortcut around documenting which sources you used and why. Pick your method, show your work, and say so in the file.
4. Why not just average everything and move on?
Because averaging hides the disagreement instead of resolving it. If four sources conflict for a reason, the reason matters. A borrower whose overtime dropped because he changed units is a different story than a borrower whose overtime dropped because the work dried up. The average looks identical. The risk does not.
5. What does “unresolved” look like on a real loan file?
It looks like a note that says: these two sources disagree, here is the gap, here is what we need to close it. Then a condition that asks for exactly that item. It is not a decline. It is an honest open question with a named next step, which is far more useful to everyone than a confident number nobody can trace.
6. What is provenance, and why should a loan officer care about it?
Provenance just means a fact stays connected to where it came from. In practice it is the difference between a file that says “income: $6,420” and a file that says “income: $6,420, from the 24-month average of base plus overtime per the 8/15 VOE and the 2025 W-2.” The second one survives a second look. The first one does not.
7. Why does the date on a document matter so much?
Because information from different periods is not interchangeable. A paystub from spring and a verification from late summer describe two different moments in a borrower’s life. Treating them as one picture is how a file ends up with a number that was never true on any single day.
8. How do I document my reasoning without writing a novel?
Two or three sentences. Source, method, and anything you deliberately excluded. If it takes you longer than ninety seconds, you are overthinking it. Build a short template and reuse it on every file until it becomes muscle memory.
9. As a Realtor, what is the equivalent of this in my work?
Value and property facts. Square footage, acreage, year built, flood zone, taxes. Your sources for those routinely disagree, and your client is making a six-figure decision on top of them. Name your source every time you give a number, and say plainly when a number is unconfirmed.
10. Does this mean AI is going to replace underwriters or loan officers?
Not the part of the job that matters most. AI is very good at reading and comparing. Deciding whether the evidence is actually sufficient, and being willing to say “not yet,” is judgment. That is the part you get paid for, and it is the part worth getting better at this year.
Source: Gerald Green, “AI can read the mortgage file, but who decides which facts are true?” HousingWire, September 16, 2026.
Kate Deiboldt
Senior Mortgage Advisor, VanDyk Mortgage Corporation
NMLS #18487 | Company NMLS #3035
Kate@VanDykMortgage.com | (931) 980-9764
Licensed in TN, KY, AL, FL, GA, TX, IL | Equal Housing Lender
Leave a Reply