An AI underwriting analyst learns your firm from the corrections you send it, not from a settings page. You reply to a memo with "next time…", "we never…" or "our threshold is…". The correction becomes a rule kept for your firm, every later review checks it, and the next memo says "as you asked" where it applied it.
Updated 22 October 2026. Published by Titleman.
The loop in five steps
- Forward. You send a deal the way you already do: the offering memorandum, the rent roll, the T-12, the broker's email.
- Review. The analyst returns a memo with its findings, each tied to the page or cell it came from.
- Correct. You reply in plain words. "Next time, use our vacancy, not the broker's." "We never count parking income on this asset type."
- Rule. The correction is written down as a rule that belongs to your firm.
- Next review. The next memo applies the rule and says where: "as you asked, vacancy is at your figure."
That is the whole mechanism. The more deals you review together, the faster it works the way you do. The page how your analyst learns shows the same loop with the screens.
What a correction turns into
| You reply | Rule kept for your firm | What the next memo does |
|---|---|---|
| "Next time, underwrite vacancy at our number, not the broker's." | Vacancy: the firm's own figure | Uses your figure and notes the broker's next to it |
| "We never buy with less than five years left on the anchor lease." | Red line on anchor lease term | Flags any deal that crosses it on the first page |
| "Our threshold is a 6.5% going-in cap rate." | Going-in cap rate threshold | Shows the deal against your threshold, not a market average |
| "Stop listing comps older than two years." | Comparable sales window | Leaves older sales out, or marks why one was kept |
The thresholds in this table are examples of what a client might say, not recommendations.
A real exchange: it shows the change before it makes it
Learning your style only helps if the analyst is careful with your numbers. Here are excerpts from one conversation on our own test account in October 2026, on an Excel model of a small office building:
- The user wrote: "Raise the monthly rent of suite 102 to $4,200." The reply: "This would change 1 cell and recalculate 9. Nothing is written until you apply it."
- The user wrote: "Set the cap rate to 6.25%." The reply started: "Nothing would change. The cap rate (Operating!B12) is a formula (NOI ÷ Value), not an input cell". In this conversation, it refused the change.
- The user asked: "What is the NOI now, and which cells does it come from?" The reply: "NOI: 216588 (Operating!B10). Before your last change it was 209676."
These excerpts show a change preview and an answer tied to a workbook cell. They are one test conversation, not a guarantee that every request is handled correctly. Check proposed changes and calculations before relying on them.
What it learns first
From the first replies, in roughly this order:
- Red lines. What your firm never does. These are the cheapest to state and the most costly to miss.
- Thresholds. Your going-in cap rate, your coverage minimum, your hold period.
- Assumptions. Vacancy, rent growth, expense loads, reserves: your numbers, not the broker's.
- Sources you trust. Which brokers' figures you take, which you always check.
- Format. What goes on page one of the memo for your committee.
In week one it knows the market but not your firm. Over the first months it picks up your assumptions, thresholds and the brokers you trust, until it reads a deal the way your team would. How fast that happens depends on how many deals you send and how much you correct. We measure it on your deals as the match with your team's decisions, and share it weekly; see measuring an AI analyst week by week.
Your rules stay yours
Your files and your rules are used for your firm's reviews only. They are not put into shared models without your consent. Another firm's analyst never sees your red lines.
When a correction does not stick
A correction that is not applied next time is an error on our side. It is fixed and the fix is reported to you, not explained away as a learning curve.
What you can do to make it learn faster
- Send everything on a deal, including the broker's emails.
- Correct by reply as soon as you read the memo, in your own words.
- Say your red lines in the first week.
- Bring a second person from your team into the replies, so it learns more than one reader.
The easiest way to see the loop is a short probation on live deals: hire Titleman on a 7-day probationary period.
FAQ
How does an AI underwriting analyst learn a firm's style?
From your corrections. You reply to a memo with "next time…", "we never…" or "our threshold is…". The correction is written down as a rule kept for your firm, every later review checks it, and the next memo says where it applied it.
Is my firm's data used to train a shared model?
No. Your files and your rules are used for your firm's reviews only and are not put into shared models without your consent.
Do I need to write prompts or settings to teach it?
No. You correct it the way you would correct a junior analyst, in plain words, by replying to the memo.
What happens if a correction is not applied next time?
It is treated as an error on our side and fixed, not explained away as a learning curve.
How long until it works the way my team does?
It depends on how many deals you send and how much you correct. In week one it knows the market but not you; over the first months it picks up your assumptions, thresholds and red lines. We measure the match with your team's decisions on your own deals and share it weekly.
Read next: hiring an AI analyst for real estate: the first 7 days, AI agents vs software for CRE firms and AI real estate underwriting software. To try the loop on a deal you know, request a demo.