AI Underwriting Red Flag Report for Lenders: Fixing Blank Rehab Items and Mixed Comps

An AI underwriting report is only useful if the underwriter can act on it in the first minute. At one private lender, the tool was scoring deals unfairly, mixing two kinds of comps, and burying the points that needed a decision under everything else it had to say.

The client is a leadership team at a private real estate lender. A new stakeholder joined the review, so we walked through the system as a first-time user would see it. If you run lending for investors, our page for mortgage loan officers and private lenders covers the wider work.

The situation

We walked the new stakeholder through the AI underwriting system and used a live deal to calibrate three things at once: the scoring, the comps logic and the design of the output. Reviewing on a real deal is the fastest way to find faults, because the lender's team already knows what a good answer looks like.

Three problems came out of that review.

What was broken

The review turned up three faults:

• Blank rehab items were penalized. When a line in the rehab scope was empty, the scoring treated it as a defect and marked the deal down for something that was not wrong.
• Sale comps and rental comps were mixed. A flip should be judged against sales, a rental against rents, and a single blended set blurs both.
• The output buried the action. The underwriter had to read through the analysis to find what needed attention.

What we built

We agreed three changes, and a pilot to test them.

• Separate comp tracks. We duplicate the scheduled comps into two tracks for each market, one for sales and one for rentals.
• A fair rule for blanks. We update the knowledge base so that blank rehab items stop counting against a deal.
• A red-flag-first output. The report will lead with a list of red flags, followed by a drafted email to the borrower with the questions the underwriter needs answered. The underwriter reads the problems first and can send the questions without writing them from scratch.
• A pilot set. The lender curates one flip and one rental per market, with the borrower tiers attached. We run them over the weekend for a review early the next week.
• An appraisal feedback loop. The lender sends the appraisal for the demo deal, so Claude can compare its own reasoning to the actual appraised values and surface the assumptions it got wrong.

What changed

The summary of this call reports no results yet, since the pilot had not run. What changed was the design of the system. The agreed design gives the tool a defined rule for blank inputs, comps that match the type of deal, and a report built around the underwriter's next action.

The appraisal step deserves attention. Checking the tool's reasoning against a known appraised value turns every demo deal into a test of its assumptions. It is how a team finds out why the tool was wrong, not just that it was.

Design your own underwriting report

If you use AI to screen loans, review the output the way the person who reads it does. Put the items that need a decision at the top. Have the tool draft the borrower questions so the follow-up is one edit and a send. Check that an empty field is handled as missing information, not as a failure. And keep the comps for sales and rentals apart, because they answer different questions.

Frequently asked questions

Why should an AI underwriting report lead with red flags?

Because the underwriter needs to act, not read. Putting the red flags first, followed by a drafted email of borrower questions, means the first thing on the page is what needs attention, and the follow-up to the borrower takes one review instead of a rewrite.

Why were blank rehab items hurting the score?

The scoring treated an empty line in the rehab scope as a problem. A blank item is not necessarily a problem, so the deal was marked down unfairly. The fix was to update the knowledge base so blank items stop being penalized.

Why separate sale comps from rental comps?

They measure different things. Sale comps support a flip's resale value, and rental comps support a rental's income. Mixing them blurs both, so we split the scheduled comps into a sale track and a rental track for each market.

What does a pilot of one flip and one rental per market test?

It covers both deal types in every market without overloading the review. Running them over a weekend with borrower tiers attached gives the team a short, representative set to check before the changes go wider.

If your AI tool produces long reports your team skims, book a 30-minute AI follow-up and reporting audit. We look at what your underwriters need first and how the output should be ordered.