RReshape.
← All articles

AI Underwriting Scores Too Conservative: What One Lender's Diagnosis Found

By Alex Regueiro · November 16, 2026 · 6 min read

A real estate lender's AI underwriting screen was scoring deals too conservatively, and the fault could sit in the guidelines, the documents or the screen itself. A score that runs low is easy to notice and hard to explain.

The fix involves the lender's principal and the person who provides its underwriting logic. If you run a lending team and your own screen feels harsher than your credit committee, our page for mortgage loan officers shows the kind of systems work we do. This article covers the diagnosis, because the diagnosis is where the useful part is.

The situation

The lender had an AI underwriting screen that scores incoming deals. On the call we looked at what it was producing, and the scores were conservative. That is the plain version of the complaint, and it is not yet a diagnosis.

A conservative score has several possible causes. The guidelines inside the screen can be too tight. The screen can misread the documents it is handed. Or the logic your team actually uses can be missing from the screen altogether. Each cause has a different fix, so the first job was to work out which one we had.

What was broken

The principal identified three problems. The screen was not reading credit scores correctly, it was not reading after-repair values correctly, and it did not handle refinance deals correctly. All three are inputs to a score, so an error in any of them can distort the result.

  • Credit scores: the screen was not reading them correctly.
  • After-repair values: the same problem, and it matters for any deal priced off a finished value.
  • Refinance deals: the screen did not handle them correctly.
  • Missing logic: a prior working session with the person who writes the underwriting logic had produced new logic on micro-neighborhoods and rent comps, and none of it had been added to the screen's knowledge base.

What we set up

The fix has two parts, and they are separate. One is a document parsing problem. The other is a knowledge problem. Mixing them is how a team ends up retuning guidelines to compensate for a screen that simply cannot see a credit score.

We set a working session to build the knowledge base. The principal brings guideline insights and feedback from rescoring deals. The person who writes the logic brings the underwriting rules, including the micro-neighborhood and rent comp logic that was never captured. Before that session we prepare a question set.

On our side the work is to fix how documents are parsed and how refinances are handled, then encode the guidelines the team gives us.

What changed

No rescoring result was reported on this call, so there is no before and after score to show. What changed is the shape of the problem. A vague complaint that the scores were too conservative became a short list: three reading and handling defects in the screen and one gap in its knowledge base.

That matters because it tells the team what to test first when the fix is in. If credit scores, values and refinances now read correctly and the scores are still low, the remaining cause is the guidelines, and that is a conversation about policy, not about software.

Test your own screen in an afternoon

This part is advice, not a description of the lender. Pick a handful of deals your team already decided on and run them through your screen. For each one, check three things before you look at the final score.

First, confirm the screen read every input the score depends on, such as credit, value and loan purpose. Second, confirm that each deal type you fund has its own logic and is not forced through a single path. Third, ask your underwriter what rules they apply that have never been written down. Those unwritten rules are the usual reason a screen disagrees with the people who know the business.

Frequently asked questions

Why would an AI underwriting screen score deals too conservatively?

Usually one of three reasons: the guidelines inside it are tighter than your real ones, it misreads inputs such as credit or value, or it lacks logic your underwriter applies from experience. In this case the team found misreads and missing logic.

How do I tell a parsing problem from a guideline problem?

Check the inputs first. If the screen reads credit, value and loan purpose correctly and the score still looks low, the guidelines are the cause. If an input is wrong or skipped, fix the parsing before touching any guideline, or you will tune rules to hide a reading error.

What is a knowledge-base build session for an underwriting screen?

It is a working session where the person who owns the underwriting logic walks through the rules while we encode them, guided by a prepared question set. The aim is to capture rules that live only in someone's head, such as how a micro-neighborhood or a rent comp changes a decision.

Do refinance deals need separate logic in an AI underwriting screen?

Often yes. A refinance has a different purpose and different inputs than a purchase or a rehab loan, so one shared path can score it wrongly. Here the refinance handling was one of the named defects.

If your underwriting screen disagrees with your team and nobody can say why, book a 30-minute AI follow-up and reporting audit. We look at what it reads, what it skips and which of your rules it has never been given.