Open any lender’s LOS and the pipeline looks under control: statuses green, documents uploaded, tasks assigned. The work that actually decides whether a loan is saleable, recalculating the income, reading the appraisal, checking the file against the guideline, still happens in an underwriter’s head and a spreadsheet.
Lenders have spent a decade digitizing the front end. Borrowers apply online, upload documents, e-sign disclosures. The systems underneath (LOS, OCR, AUS integrations, RPA bots, dashboards) collect and move data. They do not interpret policy, reconcile a contradiction, or make a defensible call. That still depends on people, one file at a time.
The old way
A processor assembles the file. An underwriter reads the returns, keys numbers into a cash-flow worksheet, cross-checks the DU or LPA findings, and clears conditions from memory of the Selling Guide. A QC reviewer later re-does a sample of that work after the loan has closed, when a defect is expensive to fix.
The agentic way
An agent reads the file the moment it lands. It recalculates qualifying income and shows the method. It flags the tradeline that appeared after application and was never re-run against DTI. It checks each condition against the current guideline and cites the section. The underwriter still owns the decision, but reaches it with the arithmetic done and the exceptions already surfaced.
The point is not speed for its own sake. Without this, the defects that trigger a repurchase demand sit undetected until post-close QC finds them, or until the investor does.
What “interpret policy” actually requires
The hard part is judgment on text that changes constantly. That is what guideline intelligence does: retrieval over the Fannie Mae and Freddie Mac Selling Guides with the exact section pinned and every citation validated, so an answer is defensible rather than plausible.
Mortna will run 50 of your closed loans and show the defects your team missed, before you commit to anything.