Claude Loan Underwriting Automation: From Borrower Email to Score in One CRM Field

A private real estate lender had its underwriting skills sitting in one staffer's Claude account, and every deal still cost a person three to four hours of copy-paste. The setup around them belonged to one person's login.

We worked with the operations and credit executive and the head of sales to reset it so the company owns the process. This is the design we agreed on and where it stands. Lenders and mortgage loan officers who want AI in underwriting without giving up approval control may find the structure useful.

The situation

Deals were emailed to one staffer. The underwriting skills ran in that staffer's Claude account, and a person copied the data across by hand. Intake went through a personal inbox, and no calls were being recorded.

The question on this call was where the skills should live and how a deal should reach them. The plan also calls for 5 to 8 hours of interviews with experts to define the skills the company account will hold.

What was broken

Four problems stacked up.

• Ownership. The skills lived in one staffer's account, not the company's.
• Manual transfer. Each deal took 3 to 4 hours of copy-paste before any analysis started.
• Personal intake. Deals arrived through a personal inbox rather than a shared one.
• No call record. Calls were not recorded, which is why call recording with summaries is part of the fix.

What we designed

We reset the system around a single admin Claude account owned by the company and connected to the CRM. The flow, step by step:

• A borrower email updates the borrower's record in the CRM.
• A mailbox watcher captures the deal spreadsheet and the purchase agreement into one custom field on the record.
• That capture triggers Claude underwriting, which writes a score from 1 to 100 along with red flags and a yes or no rationale into the field below it.
• A task notifies the responsible person to approve or decline.
• A high score, about 70 and up, triggers the check in the loan system. Zapier bridges the steps where no native connector exists.
• A human approval loop stays in place throughout. Claude scores and explains. A person decides.

How we will test it, and what surrounds it

The first test is one live deal, run so that no existing records are touched. That keeps the first proof small and the risk contained.

Around the scoring flow, the design adds three supports: a shared deals inbox to replace the personal one, call recording with summaries that become a knowledge base, and an exceptions tracker that feeds it.

What changed

No result is reported yet. A demo of the flow was promised for the following week, so what exists is an agreed design and an ownership decision.

The decision that matters most is ownership. Under the design, the skills belong to the company account instead of to one login, intake moves to a shared inbox, and the score lands in a field on the borrower's record where the team can see it and act on it.

Design the approval gate first

If you build something similar, decide who approves before you decide what scores. Keep the score and the rationale in separate fields so a reviewer can disagree with the reasoning, and keep the threshold for triggering the next check a setting you can change. Test on one deal before you let it touch the rest of your pipeline.

Frequently asked questions

Does Claude approve the loan in this design?

No. Claude writes a score from 1 to 100 plus red flags and a yes or no rationale. A task then notifies the responsible person to approve or decline, and the human approval loop stays in place throughout.

Why put the score in a CRM field?

The field sits on the borrower's record next to the deal materials, so the team sees the score where they already work, and the score can trigger the task and the next check without anyone copying data by hand.

Why use one company-owned admin Claude account?

When the skills live in one staffer's account, the company does not own its own process and the work depends on one login. An admin account owned by the company keeps the skills with the business.

How long does it take to define underwriting skills?

The plan here budgets 5 to 8 hours of interviews with experts to define the skills. Written rules are what let someone other than the expert repeat the analysis, so that time is worth protecting.

If your underwriting knowledge lives in one person's head or one person's login, book a 30-minute AI follow-up and reporting audit. We map where deals enter, where data gets copied by hand and where a scored, human-approved flow could replace it.