Titleman has not measured an accuracy rate across its runs, so it quotes none, and a page that did would be inventing a figure. AI underwriting is accurate enough for an investment committee when the committee treats it as one checked input, and checks it the way it checks any underwriting: where each input came from, what is assumed, how confident the preparer is and why, and what happens when a figure is wrong.
Prepared by Titleman AI, on behalf of the Titleman team. Figures are from finished Titleman runs and measurements published on the pages linked here, made on Titleman QA accounts or read from production, not from client files. A run is a draft for a professional to check, not an appraisal, investment advice or a credit decision.
What a committee can check
- Observed, assumed or computed. In one underwriting workbook every input was classified as sourced from a record, assumed, or computed from the other two. Reviewers could separate what the model observed from what it guessed, line by line.
- Confidence with reasons. One broker opinion run on a Tampa industrial warehouse scored itself 45 out of 100 with a High risk rating. It had found two comparable sales, hit a zoning conflict it left unresolved and lacked occupancy data. It also named what would raise the score. Below three comparable sales, a point value becomes a range.
- Independent recalculation. The corrected 416-unit workbook was recalculated twice to the cent, once against the file with a separate formula engine and once from a second implementation of the model.
- Checks inside the file. Four automated checks run in the workbook: going-in yield, exit cap against entry cap, DSCR ceiling and operating expense ratio.
A visible correction
The first draft of that 416-unit underwriting returned a levered IRR of 53%. The formulas were right; the purchase price cell held the assessed value, a tax figure, not a price. The corrected version derives price from Year 1 income and returns 8.66%. We publish the error because a committee should ask every preparer whether it has ever disclosed one. See why runs disagree.
What a run counts
On a deal package, a Titleman review opens with how many items it found and how many it checked. One six-file package: 1,788 figures, formulas and statements checked, 0 left unchecked, 6 red flags and 46 more items marked to review. A short, specific list is what a person who signs can act on.
What it does not replace
One run is not evidence that the process is error-proof. AI underwriting does not replace the committee's judgment, a licensed appraisal, a title search, a site inspection or legal review. Do not extend the discipline shown on one workbook into a claim about every output.
To test it, send a deal your committee has already seen and compare the two.