An underwriter needs to know whether a self-employed borrower can qualify on one year of tax returns instead of two. The answer is in the Selling Guide. So is a retired overlay, a generic rate sheet, and a thousand-plus other pages. Keyword search hands back all of them and lets the underwriter sort it out.
For two decades the lender intranet has been where documents go to die: a SharePoint maze that takes ten clicks to reach the current guideline and a library-science degree to be sure it is the current one.
Keyword search fails because the questions are contextual, not lexical. “One year of returns, self-employed” returns:
- the Selling Guide (well over a thousand pages)
- a lender overlay retired last quarter
- a generic FAQ on income documentation
The underwriter opens the PDF, searches for “self-employed,” checks whether the definition moved, cross-references the latest agency announcement, and hopes nothing changed yesterday.
A reasoning engine, not a retrieval engine
A grounded model reads the guideline, the announcements, and the overlays together and answers the question asked:
“Per Selling Guide B3-3.5-01, one year of personal and business returns is permitted when the business has been in existence for five years and the borrower has held at least 25 percent ownership for the past five consecutive years. This borrower shows three years of ownership, so two years of returns are required. Route it to an underwriting exception or document the second year.”
That answer is defensible because it names its source. A generic “chat with your PDF” tool guesses. A grounded one pins the exact section and validates the citation before it shows you anything.
Permission-aware retrieval is the hard part
You cannot drop every document into one index and let everyone query it. Retrieval has to check entitlements first: a reviewer scoped to one lender’s pipeline should never see another’s files, and borrower NPI should never surface to someone outside its need-to-know. The search is filtered by identity before the model sees a chunk.
From answer to action
The next step is read-write. Instead of only finding the rule, the agent flags the missing verification of employment, checks whether the file meets clear-to-close conditions, and drafts the borrower request for the missing bank-statement page.
The folder-search era is over. Lenders who replace it with a cited, permission-aware knowledge layer spend less time hunting for the rule and more time applying it.