Training an AI Deal Analyst on Your Underwriting Rules: A Two-Example Start

If the rules you use to price a deal live only in your head, nobody else can run the analysis and an AI cannot either. That was the starting point for a real estate team lead in Northeast Ohio who sells investment deals to a buyer list. There was also no central pricing sheet for contractors and materials.

This is the second part of a series about building an AI deal analyst for that team. If you work with real estate agents and teams that evaluate investment deals, the setup below shows how rules get out of a person's head and into a system.

The situation

The team lead already had material that described the method: recorded training sessions, past deal analyses, and a list of buyer tags in the real estate CRM. What was missing was a place where an AI could learn the method and be checked against it.

The goal of the call was to create that place and test whether it could reproduce the team lead's thinking on a real deal.

What was broken

Two problems sat underneath every manual analysis.

First, the underwriting rules were unwritten. Second, there was no internal pricing sheet for contractors and materials, so there was no shared source for what repairs should cost. An analysis that depends on a person's memory cannot be handed to an assistant or to software.

What we built

We built a dedicated deal analysis project in an AI assistant and loaded it with the team's own material.

• Sources: the training transcripts, sample analyses, the buyer tag list and investment calculator references, organized in a shared document with labeled tabs.
• Calibration: the underwriting logic was tuned through multiple-choice questions and answers, instead of a long written explanation.
• Deal data: filtered MLS searches exported as CSV files for the AI to read. The AI's cloud browser could not sign in to the MLS directly, so exports were the working route.
• Upkeep: a background audit of the folder of past analyses updates the project's instructions, so the rules follow the team's real decisions.
• Shared prompts: canonical prompts and templates saved in the project, so everyone starts from the same ones.
• Pricing: the team lead's follow-up answers about contractors and materials feed a new spreadsheet in Google Sheets, which starts the missing pricing sheet.

What changed

The AI closely matched the team lead's underwriting rules from just two examples. That is an encouraging early sign, and we want to be careful with it. Calibration was not finished when the call ended, and two examples are not a track record.

The other change is structural. The rules now exist in writing and in a project that can be corrected, and the pricing answers are being collected in a sheet. Both can be reviewed by someone other than the person who wrote them.

What to take from this

This is general advice, not a description of the client's account. If you want an AI to follow your underwriting, give it your real examples and calibrate it with questions. Multiple-choice questions are faster to answer than essays, and they surface the rules you did not know you were applying.

Test it on deals you have already priced. Compare its answer with yours before you trust it on a new one. And start the pricing sheet now, because every estimate you make from memory is data you could be saving.

Frequently asked questions

Can an AI learn my real estate underwriting rules?

It can learn to follow rules you show it. Here the project held transcripts, sample analyses and a Q&A calibration. The AI closely matched the team lead's rules from just two examples, though calibration was not finished, so it still needs checking.

Why use multiple-choice questions to calibrate the AI?

Questions with fixed options are quick to answer and they force a clear decision, and each answer becomes a written rule. It is an easier way to capture logic than asking someone to describe their whole method from scratch.

Why export MLS searches as CSV files?

The AI's cloud browser could not sign in to the MLS directly in this setup. Filtered searches exported as CSV files gave it the same listing data without needing that login.

What is the pricing sheet for?

There was no central source for contractor and material costs. The team lead's follow-up answers go into a new spreadsheet in Google Sheets so repair estimates have one shared reference instead of living in memory.

If your underwriting or estimating rules exist only in your head, book a 30-minute AI follow-up and reporting audit. We show you how to get them written down and tested against deals you have already priced.