Ask an AI “is this self-employed borrower risky?” and you get a confident paragraph of nothing. Ask it the way a senior underwriter would, and you get a figure you can defend to a QC reviewer.
Role, task, constraint
A vague prompt gets a vague answer. Structure the request the way you structure a file review.
Weak: “Is this borrower’s income okay?”
Strong: “Act as a conservative underwriter. Analyze this borrower’s two years of personal and business returns. Calculate qualifying income using the Fannie Mae Form 1084 cash-flow method. Flag any year where income falls more than 20 percent from the prior year, and tell me which decline I have to document.”
The role sets the posture. The task names the method. The constraint tells it what to surface.
Give it your add-backs as examples
Generic instructions produce generic spreads. Show the model how your shop treats the hard lines: depreciation on Schedule C, non-recurring income, business use of home. Three to five worked examples of your own conventions move accuracy more than another paragraph of instruction, because the model copies a pattern instead of inventing one.
Make it show its work
If it returns a DSCR of 1.20 on an investor (DSCR) loan, make it print the formula and the exact numbers it used: qualifying rent over PITIA. A figure with no method is a figure you cannot sign, and cannot explain if the loan is later pulled for QC.
The point is not to replace the underwriter’s judgment. It is to get the arithmetic done and shown, so the human spends the hour on the decision instead of the data entry.