AI Underwriting for Private Real-Estate Lenders: What One Live Test Deal Changed

A private lender's underwriting started with retyping each application into a spreadsheet, pulling credit and tax data by hand, and judging rehab scopes and after-repair value from experience. It is slow, and it rests on individual judgment.

The lender makes short-term real estate loans to investors across several markets, the kind of work that mortgage loan officers and private lending teams share. We reviewed the Claude underwriting sequence we had built for it, and ran the first live test deal in front of the lender's head of operations and head of originations.

The situation

The lender's process was manual end to end. The CRM was disorganized, with many systems that did not talk to each other. A previous AI consultant had optimized the wrong things, so this build started from underwriting itself.

That shaped the scope. The team wanted an underwriting agent first, tested on deals they had already closed. Changes to how borrowers submit loans were held back. Borrowers send their files by email today, and that waits until the underwriting itself proves useful.

What was broken

Three gaps stood out. Rehab scopes had no scoring method, so reasonableness came from whoever read the file. Market knowledge was thin. And it did not include house age or builder history, which bear on plumbing, electrical and structural risk in older homes.

What we built

• A Claude project with skill files and a per-market underwriting scoring engine.
• Market knowledge bases for every market the lender serves. We ran the same deep-research prompt in three AI tools, then distilled the results to one page per market and tied them to the lender's market-share data through a connector.
• A first test deal run live. Claude approved or flagged each line of the rehab scope, explained its logic for each call and asked 3 to 5 confirming questions at the end.
• A test method that uses the lender's existing spreadsheet scope and purchase contract as the input, so no new borrower form is needed. Calibration starts with 10 completed-deal scopes with their appraisals, then moves to about 10 live deals.

What one feedback round changed

The most useful part of the call was the first round of live feedback. The lender's team read the output and corrected it, and Claude restated the scoring logic. Three things changed.

• Investor crews are no longer scored against retail contractor rates. An experienced borrower with their own crew builds for less, so the borrower's experience tier now sets how skeptical the scoring is of a low line item.
• The appraiser photo checklist now asks for finish quality, not just presence. A kitchen with cabinets installed can mean stock units or salvaged ones, and the appraisal treats them differently.
• House-age and builder-history knowledge is added to each market's knowledge base, with a second round of deep research to cover it.

What changed

The record states no results. The test deals were set for that weekend, so the claim here is about the system. The underwriting now lives in a project that can be corrected by the people who own the credit decision, and a single round of corrections already moved the scoring in three ways.

The team also agreed on a habit worth copying: label every idea as a future idea or a current build. Closing the loop by scoring the agent against appraisal photos and appraisal results was named as a future build, not part of this one.

What to take from this

This section is advice for any lender. Test on closed loans first, with a human making the credit decision, and expect the first run to be wrong in useful ways. Ask the model to explain its logic and to ask you questions. The questions show you which assumptions it is guessing.

Keep intake unchanged until the underwriting earns trust. A smarter scoring step is a smaller change to sell inside a company than a new process for every borrower.

Frequently asked questions

Can Claude underwrite rehab loans for a private lender?

It can review a rehab scope line by line, approve or flag each item and explain its reasoning, as it did on our live test deal. A person still makes the credit decision. We treat the output as a faster, consistent first review that the team can correct.

How do you calibrate an AI underwriter?

Start with completed deals where you know the outcome. Here that meant 10 rehab scopes with their completed appraisals, followed by about 10 live deals. Each correction from the team becomes a change to the scoring rules and the market knowledge.

Why build market knowledge bases for each market?

Costs, housing stock and risks differ by market. A one-page base per market gives the scoring engine local context for costs and risks. House age and builder history were added after the first feedback round.

Do borrowers need a new form to use this?

Not in this build. The test uses the lender's existing spreadsheet scope and purchase contract as the input, so borrowers see no change. Intake automation is a later decision, once the underwriting has proven itself.

If your underwriting still starts with retyping files and judging scopes by memory, book a 30-minute AI follow-up and reporting audit. We look at your intake and review steps and show you where a tested first pass fits.