Testing AI tools for CRE loan underwriting and monitoring

What the CRE suite covers

Our CRE suite is small and new: 8 tasks, of which 3 are clean cases where nothing is wrong and the right answer is to say so. Each task is a fictional property and loan, with the relevant clauses written out. The tool gets the documents; it never sees the expected answers.

• DSCR on underwritten cash flow, after replacement reserves and any vacancy floor the agreement imposes.
• Interest-only periods that have ended. Once amortization starts, debt service rises and a loan that passed can fail.
• Debt yield on the current balance, as the agreement defines it, rather than on the original loan amount.
• Maximum loan sizing across the LTV, DSCR and debt-yield constraints, naming the one that binds.
• Lease rollover within a 12-month window, including a lease that expires exactly on the boundary date.

We score every answer on the same four questions as our private credit suite: was the problem detected, was it measured correctly, was it escalated, and was the right clause cited. Each task runs three times so we can see whether the tool gives the same answer every time.

How a CRE test works

The same as our private credit pilot. We agree the configuration to be tested, run it through the suite three times, and send a private scorecard showing every expected answer, the tool's answer and the reason for each miss. The fee is fixed and quoted per scope, the same whatever the result.

The suite is synthetic and cleaner than real loan files, with no scans or OCR noise. It measures reasoning and calculation on stated terms. A scorecard describes one configuration at one point in time; it is evidence, not a certification.