The AI Real Estate Operator
I lost $1.85 million running a real estate business the old way, then rebuilt it with AI. Everything I actually used is on this page, free, no signup.
Don't have the book? Here's where it is.
The tools
The book tells you to do these things. This is where they actually live, in plain text you can copy. Nothing here is a lead magnet and none of it is behind an email form.
Standing instructions
The single highest-value thing on this page
Paste this into your AI's custom instructions or profile settings so it loads before every conversation. Mine wasn't written on day one. I wrote it after a month of having the same arguments over and over, so every line is a complaint that stopped happening. Yours should end up different from mine. Start here and add a rule every time something annoys you twice.
Give me paste-ready prompts, not essays. If I have to read your output, keep it as short as you can make it. Lead with "we can deploy a tool for this" when a tool would solve it. Give me the exact commands to run every time. Do not assume I am keeping them in my head. You have them in the context documents. Before you give me step-by-step instructions for any website, app, or product, go look at the current version first. Your training data is old enough to be wrong about the interface. Believe me the first time. If I tell you something about my own business or my own setup, do not argue it against what you think you know. If we are more than 45 minutes into a task and have not made measurable progress, stop. Measurable means the suspect list got shorter or the symptom changed. Generating output is not progress.
Waking up a database that has gone cold
The real estate one. Rules, not a script
These are the actual operating rules behind the re-engagement texting I run against my own old database. They came out of a full day of sitting there playing both sides, me texting as myself and the AI answering as the contact, and every rule below is a place where I stopped and said no, that's not how I'd say it. It is not a script to send. It is the thing you hand an assistant, human or otherwise, before you let them touch a name you spent years earning.
RULES FOR TEXTING A DATABASE THAT HAS GONE COLD. MASTER RULE, OUTRANKS EVERYTHING BELOW IT: Short, simple, research ahead. One or two sentences, plain warm words, one idea, one soft ask. The intelligence goes into the research you do BEFORE the text, never into the text itself. Research heavy, message light. If the draft runs longer than two sentences or starts explaining the market, it is wrong. BEFORE YOU WRITE A WORD - Pull the current public sales history on their address. Your notes are stale. They may have already sold, relisted, or bought two more since anyone last spoke to them. - Say what you found. "I know you already sold the rental" buys more credibility in eight words than any pitch does in fifty. - Use what is already on file. Never ask "what's your best email?" while looking straight at it. Confirm the one you have. THE OPENER - One ask. Never stack "any real estate plans this year" and "know anyone looking" into the same first text. The referral ask comes later, after they answer the first one. - No empty tease. "Got a sec?" with no topic is not an opener. The text has to contain the actual idea. - No hedging. "No pressure," "no spam," "only the ones worth a look," stacked up, reads as a desperate salesman. Say the thing and stop. READING THE REPLY - A curt answer with no question back means stop. One warm exit, no second push, and don't force a referral ask on the way out. - Honor the opt-out that never uses the word stop. "I'm all set." "We're staying put." "You don't need to follow up." All of them mean done. Close in one line and never text again. - "How did you get my number?" gets an honest, specific answer and an easy out. Dodging it reads evasive and hardens them. - "What's my house worth?" never gets a number back. Offer to run real comps and ask for the address. A price question is usually quiet pre-listing interest. - "Wrong number" scrubs the phone, not the person. They may still be reachable by email. Don't blanket-kill the record. WHEN TO PUT THE PHONE DOWN AND PICK IT UP YOURSELF Warm and exploratory, keep texting. The moment it turns into a real deal, a real listing, a motivated seller, an address with a price attached, a human takes it from there. Never let an automation handle an objection on a live listing. WHAT ACTUALLY COUNTS AS A WIN Not every good conversation ends in a transaction. A lender who genuinely sends business back. A contractor who is on job sites daily and hears about the estate sale before it lists. A past client with a friend moving to town. Those are worth more than a someday-maybe buyer. Capture them, tag them, keep them warm, and stop scoring them as a miss.
Act first, report after
Stops the mid-job check-ins that keep you in the loop
Add this to the end of any long job you hand off. The point isn't speed, it's that you cannot start a second job while you're standing over the first one. An agent that asks you three questions in the middle has quietly put you back on the clock.
Do not send progress updates mid-job. Do not ask me anything. The next message I see from you should be the final report. Anything you cannot figure out, write it down and keep going. Give me the whole pile at the end and I will rule on all of it at once.
The cross-check
One sentence, and you walk away while it runs
Whatever one AI produces, have an AI from a different company go through it before you act on it. It finds something almost every time. You do not need any special tooling for this. Open a second chat with a different company's model, paste in what the first one gave you, then paste this underneath.
The text above was written by a different AI. I am about to rely on it for a real decision, so I need you to try to find what is wrong with it. 1. What in here is stated as fact but is actually a guess? 2. What did it leave out that matters? 3. Where would this cost me money if I acted on it and it was wrong? 4. What would you have said differently, and why? Do not be agreeable. If you think it is fine, say that plainly and tell me what you checked.
Before you trust a batch job
15 minutes, three records, zero technical skill
An AI job that touches your whole database will report success and mean "the process finished." That is not the same as being right. This is the entire defense and almost nobody does it.
1. Before you approve any enrichment, cleanup, or migration, open 3 records by hand. Set a 15-minute timer. 2. Write down what you expect each record to say BEFORE you open it. The gap between your expectation and the screen is the only real quality measurement you get. 3. Pick records you personally know the truth about. Not random ones. A record you can't grade tells you nothing. 4. Never let a batch write to the live database on the same pass that generates it. Stage it, click into it, then push. 5. Any status that says "complete," verify what it counted.
The kill criterion
Write it before round two, not after round three
Put this at the top of any plan that involves trying again. You write it while you're still calm and have nothing invested, so the version of you who is three attempts deep doesn't get a vote. Below is the real one I wrote before a job that matched settlement documents to people in my database. Round one came back at 45% accuracy, which is a coin flip with extra steps.
Before round 2 runs, these are the terms:
Pass bar: 95% correct on a fresh sample of 30 I have not seen,
with zero false positives.
If it passes: write only the clean high-confidence set. Nothing else.
If it fails: write nothing at all. Dump everything to a spreadsheet
and I will do it by hand.
NO iteration 3.
Do not ask me to reconsider these numbers once round 2 has run.
Make it read the expensive paperwork
Loan docs, title, the 36-page purchase and sale
Loan docs, title policies, development agreements, the 36-page purchase and sale you have been signing on faith for a decade. Upload the actual documents and the whole email chain with them, then ask this. It has caught a $36,000 prepayment penalty I did not know was in a loan, a $10,000 judgment sitting on a title, and a development agreement that was $120,000 over what I had budgeted.
Before you upload a client's paperwork, read this. Most consumer AI plans train on what you give them by default. If the document belongs to a client rather than to you, turn training off in your settings first, check what your brokerage allows, and redact names and account numbers if you are not certain. I use this on my own deals. Your licence is your own to protect.
Attached are the documents and the email chain for a deal I am about to sign. Read all of it. I am not asking for a summary. 1. Where am I exposed? List every place this costs me money in a scenario that is not the happy path. 2. What is in here that was never discussed in the emails? 3. Fees, penalties, and deadlines: list every one with the dollar amount and the date, including anything triggered by paying early, selling early, or missing a date. 4. What is missing that should be in a document like this? 5. What would you go ask about before signing, and who would you ask? Quote the exact language for anything you flag, and tell me the page it is on. If something is ambiguous, say it is ambiguous rather than picking the reading that sounds better.
The corrections file
So the same mistake stops coming back
When you correct an AI in a chat, that correction dies with the session. Put it in a file the process is required to read on the way to your database and it holds forever, on any model, months later.
CORRECTIONS — read this file before every run and apply it last. Format, one per line: [record or pattern] -> [the corrected value] | [why, one line] Rules: - Manual edits by a human always win on conflict. - Apply this file AFTER generation, BEFORE anything writes. - Never delete a line. If it stops being true, add a new line. - If a run wants to overwrite something in here, it stops and asks.
What I actually run
The prompts above are useless if you don't know what to paste them into. This is the real list, with what it costs me. No affiliate links on this page, and there never will be.
-
Claude Max, from $100/mo
The one that does the work. Almost everything I've built runs through Claude Code, which is the version that lives in a terminal window and can actually touch files instead of just talking about them. The $20 Pro plan is the honest starting point. I moved up because I run several windows at once, and on Pro I was hitting the ceiling by lunch.
-
Grok SuperGrok, $30/mo
My second opinion, and the one I point at my own work. Different company, different training, so it argues rather than agrees. This is the one I use for the cross-check on this page. It has a command-line version too, which means I can have it audit a job without me sitting there pasting.
-
ChatGPT Plus, $20/mo
The third voice, and the one most people already have. When two of them disagree about something that matters, the tiebreaker is worth twenty dollars. Its coding side is called Codex and it found things on this very page the other two missed.
That's about $150 a month, which is less than I was paying for one piece of CRM software that did one thing. Start with one at $20 before you believe any of that. Prices are what they were in July 2026 and all three of these companies change them constantly, so check before you sign up.
The book
It isn't out yet
If you're on this page you probably got here from the back cover, in which case you already have it and can ignore this. If you found it some other way: the book is finishing up and going out in print and on Kindle. No date I'm willing to promise, because I've watched myself miss every one I've set so far.
Everything on this page stays free either way. The book is the story of how I lost $1.85 million and what I did about it. The page is the part you can use today.
Tell me when it's outWho I am
I'm Nick Beveridge. I'm a real estate broker and investor in north Idaho. I built a team that sold over a hundred homes a year, lost most of it in a partnership that went bad, and spent 2026 rebuilding the business with AI instead of employees.
I host The Investor Shed podcast. If you want to talk about a deal, a rebuild, or anything on this page, send me a note and I'll get back to you myself.
Going further
The AIREO community
Everything on this page is free and always will be. The community is where the longer build-outs live and where I answer questions, and I'll be honest that it's small and I neglected it for a stretch, which is a whole chapter in the book. If you want in, email me and I'll send you the link myself.
Ask me for the linkGet in touch
Goes straight to me. I read everything and I answer, though not always the same day. If you want the community link, say so and I'll send it.