Every math teacher I had wrote the same three words in the margin: show your work. A right answer with nothing behind it got partial credit, because the teacher could not tell whether I understood the problem or got lucky.
Credit works the same way. A loan committee approves the reasoning behind a number as much as the number, and an examiner who opens the file two years later wants to follow that reasoning without calling anyone.
Why an answer alone fails in lending
General-purpose AI is good at producing answers. Ask it for a debt service coverage ratio and it will give you one in a few seconds. What it often cannot tell you is which page of which document each input came from, or whether the rent roll and the operating statement agreed before it did the division.
For a marketing email that is fine. For a credit decision it is a problem, because the lender owns the outcome. If the figure is wrong, a chief credit officer has to explain it to a board and a regulator, and "the system told me" will not carry that conversation.
Lenders tell us this in their own words. The questions we hear most are about accuracy, about where the numbers come from, and about who is accountable when the machine is wrong.
What showing the work looks like
When we built LenderBox, we started from that margin note. Every figure in the analysis is cited to the document and page it came from. An underwriter can open the page behind any number in the memo and check it against the tax return or rent roll.
The same rule applies to policy. Each deal is checked against the lender's own credit box, and every exception goes on the record with the reason. When two documents in the file disagree, the disagreement is shown with the pages behind it, so a person can decide which one is right.
And the decision stays with people. An underwriter reviews and approves at every stage. LenderBox drafts the credit memo and a recommendation, and your credit officers make the call.
What it does for the people in the room
The analyst stops retyping and starts checking. Review time goes to the items that need judgment.
The credit committee can test any figure in the memo quickly, and spend the meeting on whether this is a loan the institution wants.
The examiner finds a file that explains itself: the figure, its source, the policy test, the exception and who approved it.
Where the machine stops
We do not publish an accuracy percentage and we do not promise perfection. We design for near-perfect accuracy, and a machine can still make mistakes. That is the reason the work is shown. A cited figure can be checked by anyone on the credit team. An uncited one has to be rebuilt by hand before anyone can rely on it.
Three questions for any AI vendor, including us
- Can I trace every figure to the document and page it came from?
- When the file contradicts itself, what do I see?
- Who approves each step, and can the software ever approve a loan?
Clear answers to those three tell you whether your credit team can stand behind the tool. The fastest way to get them is to watch it run on one of your own files.
Bring a deal and we will show the work on your documents.
Common questions
What does "AI that shows its work" mean in lending?
Every figure in the analysis is cited to the document and page it came from, every policy exception is recorded with its reason, and a person approves each step.
Can AI approve a commercial loan?
It should not. LenderBox drafts the credit memo and a recommendation, and the lender's credit officers make the decision.
How accurate is AI underwriting?
Judge it on your own files. LenderBox is designed for near-perfect accuracy, a machine can still make mistakes, and a person stays in the loop for every judgment call.



