A lender can compare AI underwriting engines the same way it tests an analyst: hand each one a file already underwritten and check the answer. Six points separate a useful engine from a fast demo. Each is checked below against Titleman, using finished runs and not claims.
How should a lender choose AI underwriting software for CRE loans?
Score every candidate on the same file, on these six points:
- Package tie-out. Does it read the rent roll, the T-12 and the offering memorandum against each other?
- Stress. Does it restate DSCR and value at today's rates?
- The memo and its open items. Does it list assumptions and what it could not establish?
- Traceability. Does each figure point to a page or cell?
- Delivery. Where does the result land?
- Run time, as measured. What is the median, and what is the slow end?
Does it tie the package to itself?
On one 50-unit loan package, the rent roll and the T-12 agreed with each other. The offering memorandum's in-place income of $614,475 did not: it was 32% above the T-12's $466,880. Ask each vendor to find that kind of gap on your file, and to show where it came from. Our line-by-line teardown does so.
Does it stress DSCR and value at today's rates?
A coverage ratio from the marketing package is not a credit test. In a finished run on a sample 120-unit package, a $14,040,000 balance covered 0.97x on the T-12 income at a 6.5% refinance rate, against 1.23x on the memorandum's income. At a 6.5% cap rate, the same income supported about $15.8 million of value, so the value test, not coverage, set the refinance at about $10.3 million, a gap of $3,749,000. The full stress shows every row.
Does the memo list its assumptions and its gaps?
A loan-sizing run on a 700-room hotel, made without the borrower's operating statement, recorded 679 fact rows. It marked 143 as missing instead of estimating them and labeled 21 as assumed. That list is the document request for the borrower. A tool that fills every cell has hidden the same gaps. See the full run.
Can you open the source behind a figure?
Pick any number the credit committee will question and ask for its page or cell. Titleman ties each finding to the place it came from. Also ask how the vendor describes model risk. The Federal Reserve's model risk guidance is a useful prompt for that conversation.
Where does the result land?
The hotel run returned an Excel workbook and a memo in Word, Markdown and PDF. The Excel add-in is live. Email forwarding into a workspace is listed as not available yet, and we do not claim a loan-origination or CRM push, so ask every vendor, us included, to show delivery into your own stack.
How long does it take?
Measured, not promised: the 120-unit package run was delivered 49 minutes after upload, and the hotel run took 1 hour 12 minutes. Two runs are not a median. Ask for the median and the slow end over many runs. The "72 hours" on our refinance page is a suggested schedule for a book of maturing loans (request files, run the stress, decide). It is not a service-level promise, and the time we have not measured is the time your borrowers take to send files.
What does AI underwriting not replace?
It does not replace the credit decision, a licensed appraisal, a title search, a site inspection or legal review of the loan documents. Titleman returns a draft and a list of open items for a credit officer to check. A market-based screen is not a credit approval.
Run it on a loan file you already know
Request a demo and send a file your team has already underwritten. Compare the two line by line.
Prepared by Titleman AI, on behalf of the Titleman team.