What should an AI stack for a real estate business have in it?
Written for agents, brokers, lenders, contractors, home services.
The short answer
Think in layers, and define each layer as a job rather than a product. There are seven: capture, records, pipeline, transactions and documents, the assistant, connections, and the check on what leaves.
Most businesses have four of the seven, bought one at a time, and the two that are almost always missing are the records layer and the check on what leaves.
Sort before you rank by urgency. Ranking first fills the urgent list with work that should have been refused.
The most expensive finding is two tools holding the same records, because two systems holding one set of records produce two answers to one question and neither announces that it has gone stale.
Anything that sends has exactly one recipient, set at install, and it is the owner. On 2026-07-31 an automated brief containing private client information reached eighteen unintended recipients on a comparable system. The cause was never identified and no fix was ever documented.
Nobody buys an AI stack. People buy one tool at a time, on a Tuesday, because something went wrong on Monday. Five years later there are eleven subscriptions, four of them doing the same job, and no answer to the question of where a client fact is supposed to live.
This piece is the layer map we use before recommending anything to anybody. It is written as jobs rather than products, because products change every eighteen months and the jobs have not changed in twenty years. If you want the sorting method on its own, it is in the four box sort, and everything below assumes it.
What should an AI stack for a real estate business have in it?
Seven layers, each defined as a job: capture what you said and heard, hold what you know about a person, move people through stages, keep signed documents findable, run the model that reads and drafts and files, connect that model to your accounts, and check anything that leaves under your name.
Every product on the market fills one or two of these. Nothing fills all seven, and anything claiming to fill all seven is filling some of them badly.
The value of naming the layers is that it lets you notice an empty one. A gap in a layer nobody has named looks like a personal failing, and people respond to it by working harder rather than by filling the layer.
The seven layers
| Layer | The job, in one sentence | What usually fills it | What breaks when it is missing |
|---|---|---|---|
| Capture | Get what you said and heard into the system without retyping it. | Dictation, a pocket voice recorder, a meeting notetaker. | Everything the business knows stays in one person's head and in the car. Nothing downstream has anything to work from. |
| Records | Hold what you know about a person, in one place, for as long as you know them. | Plain text files in a folder, a notes app, a wiki, or nothing at all. | The same client exists in four places and nobody can say which version is current. |
| Pipeline | Move people through stages and tell you who is next. | The CRM, the platform bundled with your brokerage, a spreadsheet. | Follow up depends on memory, which means it depends on whoever is least busy that week. |
| Transactions and documents | Keep signed things findable by somebody who was not there. | A transaction management system, a cloud drive, the title company portal. | The disclosure exists and takes forty minutes to locate, usually on the day it is being asked for. |
| The assistant | Do the reading, drafting and filing that nobody has time for. | A general purpose model you talk to, with your material in front of it. | Capture piles up unread, and the layers underneath it turn into a very tidy archive nobody uses. |
| Connections | Let the assistant reach your email, calendar and drive. | Connectors, each one a key to one account. | The assistant works only from what you paste, which caps its usefulness at your patience. |
| What leaves | Check anything published, sent or signed before it goes. | Usually nothing. Sometimes a broker review on marketing only. | Private information leaves under your name and you find out from the person who received it. |
Layer by layer, and where each one usually fails
Capture. The job is to get what you said and heard into the system without retyping it. In this industry that means dictation for the twenty seconds of context after a showing, a pocket recorder for the drive home, and a notetaker for the listing appointment. Dictation is the highest return per dollar in the whole stack, and we have written about what dictation does to a working day. Notetakers carry a different set of risks and we have set those out in what an AI notetaker misses in a listing appointment. The failure when capture is missing is total and invisible: there is nothing for any other layer to work on.
Records. The job is to hold what you know about a person, in one place, for as long as you know them. This is the layer almost nobody has, because the CRM appears to cover it and does not. A record is where you keep that the wife makes the decisions, that they will not consider anything east of the river, and that you promised to call the inspector yourself. We argue the case for plain text files as the record layer, the version of it built on a folder of notes you own, and why a database tool is a view rather than a record. The failure when this layer is missing is the same client living in four places with no way to tell which one is current.
Pipeline. The job is moving people through stages and telling you who is next. This is the CRM and it is the layer everybody already has. The mistake is asking it to do the records job as well, which we take apart in whether your CRM should be where your business remembers things. A pipeline tool is built to advance a stage and report a count. It is not built to hold a sentence somebody said in a kitchen in 2019, and every attempt to make it do so ends in a notes field nobody searches.
Transactions and documents. The job is keeping signed things findable by somebody who was not in the room. Most of this is already decided for you by a brokerage platform or a title portal, which is fine. What is not decided for you is where everything else goes: the photos, the inspection video, the scanned addendum somebody sent as a phone picture. The record that fixes this is one page naming a service and a folder for each kind of file, which is the argument in where your business files really live.
The assistant. The job is doing the reading, drafting and filing. The choice that matters here is not which vendor but which size of model for which job, because running the largest model on filing work costs money every day and running the smallest model on a pricing judgment returns the first plausible answer with total confidence. We set the sorting out in which AI model for which job. The rule worth carrying: finding things is low intensity work, weighing things is high intensity work, and scheduled work defaults to the smallest model that can do it.
Connections. The job is letting the assistant reach your email, calendar and drive. A connector is a key to one specific account, and it will not tell you which account it is holding. That single property causes more confusion than anything else in the stack, and we have written it up with dates in what it means to connect your AI to your email and calendar. Two rules from that piece belong here. Read back which account an integration is bound to before you write through it. Keep always allow switched off for connector tools, so a write still needs a human click.
What leaves. The job is checking anything published, sent or signed before it goes out under your name. This layer is missing from almost every stack we look at, and it is the one where the consequences are regulatory rather than annoying. Two failure modes matter most, and both have their own piece: an automation that reports success while doing nothing, covered in how you know your automation is really working, and a machine writing down a commitment nobody made, covered in what happens when your AI writes down a decision nobody made.
In what order should you fix the layers?
Sort the whole stack first, then rank what survives by how fast it gets worse if nobody touches it. Ranking before sorting fills the urgent list with work that should have been refused.
This is the most common sequencing mistake and it feels like productivity while it is happening. Somebody lists eleven problems, ranks them by pain, and starts on the top three. Two of those three are tools that should have been cancelled rather than fixed, so the effort is spent making a removal candidate work better.
Sort first: is it in the business now, and should it be. Then rank whatever survives. Something costing you money today outranks something that decays slowly, and both outrank something that will be the same problem in six months.
The four boxes
The four boxes run across a whole stack at business level rather than product level. The fourth box is the point of the exercise and the arguments in it are the ones worth having.
There, and should be
One tool per layer, in the layers you have filled, doing the job that layer names.
- Pipeline and documents are usually already right, often because somebody else chose them for you. Leave them alone and stop trying to make either one hold what you know about a person.
- Capture is usually right and underused. The tool is installed and gets opened twice a week. That is a habit problem rather than a buying problem, and no purchase touches it.
- A tool in this box still gets a review date. A platform bundled with your brokerage can change its terms, its export options or its owner without asking you.
Missing, and should be there
The records layer, and the check on what leaves. Those two are empty in most stacks we look at.
- Write the gap as a job before naming any product. "One place where every promise made to a client is written down" is a job. A product name is an answer to a question nobody has asked yet.
- The records layer is the one that pays compounding returns, because everything else in the stack gets better when there is something true to read.
- The check on what leaves can start as a person rather than a tool. One named human reading anything that goes out under the business name is a working version of this layer.
There, and should not be
Two tools holding the same records, and always allow switched on for anything that writes.
- Two tools holding the same records is the most expensive finding in this whole exercise. Two systems holding one set of records produce two answers to one question, and neither one announces that it has gone stale.
- A tool nobody opens is cheap. A tool half the team opens is what produces the argument about whose number is right, usually in front of a client.
- Get your data out before you cancel anything. That is a separate action from cancelling and it comes first.
Missing, and correctly missing
Four capabilities we recommend against at business level, each with its reason and its cost written down.
- An automation that sends on your behalf without a human click. On 2026-07-31 an automated brief containing private client information was sent to eighteen unintended recipients on a comparable system. The cause was never identified and no fix was ever documented. The rule we run now is that anything which sends has exactly one recipient, set at install, and it is the owner. The recipient is never resolved from context, from a file, or from a message being processed. Cost of leaving this out: every send starts as a draft that lands with you, the system cannot email a client for you, and you will click more than you want to.
- A chatbot answering client questions the agent should be answering personally. The questions a buyer asks at eleven at night are the ones where the relationship is made. Handing those to a bot converts the most valuable minutes in the business into a transaction with a website. Cost of leaving this out, stated honestly: your first response is slower, and some leads will go to whoever replied first. That is a real cost and it is the correct trade for a business that runs on referrals.
- A tool that scores or ranks your contacts for you. Two problems. The first is that a score is a decision presented as a number, and nobody can explain the reasoning behind it a year later. The second is specific to this industry: a model trained on who you closed with in the past will reproduce the pattern of who you closed with in the past, and you will be the one holding it when somebody asks why. Cost of leaving this out: you do your own ranking, it takes an hour a month, and you will sometimes be wrong in ways you can at least explain.
- A system that writes decisions. A machine may record anything it hears. Only a person promotes anything to a decision, with a name and a date on it. Cost of leaving this out: somebody has to do the promoting by hand, which takes seconds and gets skipped, and unpromoted material piles up looking uncertain. That pile is the honest state of the world and it should look like that.
- Each of these four gets one written sentence in your own file, with the reason. Without the reason it gets proposed again in ninety days by somebody who was not in the room, and you will rebuild the argument from scratch.
Being honest about compliance checks
Any tool advertising a fair housing check deserves one specific question before you rely on it: which rules is it checking against.
Federal protected classes come from the Fair Housing Act. A check built into a tool runs on that federal floor unless somebody has written your state and local rules into it. State and local protected classes vary, they are broader than the federal list in many places, and they change. Source of income, sexual orientation, gender identity, age, marital status, military status and lawful occupation all appear as protected classes somewhere and not everywhere.
Never guess a local class list, and do not accept a guess from a tool. The dangerous version is a check that runs, covers only the federal floor, and reports itself as complete, because the owner reads that line and stops looking. This is the same failure as any automation reporting a check it did not run: the report is worse than silence, because silence would have sent somebody to look it up.
The workable arrangement is a check that names its scope out loud. Something reading "checked against the federal list, local rules not covered" is useful. Something reading "compliance check passed" is a hazard wearing a green tick.
Can an AI tool handle fair housing compliance for you?
Not on its own. A built in check runs on the federal floor from the Fair Housing Act unless somebody has written your state and local rules into it, and those vary by jurisdiction and change over time. A check reporting itself complete when it only covers the federal list is the dangerous version.
Ask the vendor which jurisdictions the check covers and when the list was last updated. If nobody can answer that, treat the check as a spell checker for one specific set of words rather than as compliance.
Keep your broker, your association counsel, or whoever advises you on this, in the loop for the local list. That is a human job and there is no version of this where a tool takes it over.
The three rules we would keep if we lost everything else
- 01One layer, one tool. The moment two tools hold the same records you have two answers to one question, and neither system will tell you that its copy has gone stale.
- 02Anything that sends has one recipient, set at install, and it is the owner. Never resolved from context, a file, or a message being processed. This is the rule that came out of the 2026-07-31 incident and it is the strictest thing we run.
- 03A success response is not evidence. Read the artifact back from the system that holds it. A green tick on a compliance check, a sent count on a campaign, a completed item on a transaction checklist: each of those is a claim made by a system about itself.
What this map does not do
If you read one other piece after this one, make it the four box sort, because it is the method underneath everything above and it takes twenty minutes to run on a single tool before a demo call. The rest of the library fills in one layer each, and every one of them ends with the same four boxes filled in for one product, including the fourth.
Common questions
- What are the layers of an AI stack for real estate?
- Seven, each one a job rather than a product. Capture what you said and heard. Hold what you know about a person. Move people through stages. Keep signed documents findable. Run the model that reads, drafts and files. Connect that model to your accounts. Check anything that leaves under your name. Most businesses have four of the seven.
- Which layer is missing from most real estate businesses?
- Two of them. The records layer, because the CRM appears to cover it and does not, and the check on what leaves, because nothing forces anybody to build it until something has already gone out. The records layer is where you keep what you know about a person. The CRM is where you move that person through stages.
- What is the most expensive mistake in an AI stack?
- Two tools holding the same records. Two systems holding one set of records produce two answers to one question, and neither one announces that it has gone stale. It shows up as an argument about whose number is right, usually in front of somebody who is paying you.
- Should an AI be allowed to send email for me?
- Only to one fixed recipient, set at install, and that recipient should be you. On 2026-07-31 an automated brief containing private client information reached eighteen unintended recipients on a comparable system, the cause was never identified and no fix was ever documented. The recipient should never be resolved from context, from a file, or from a message being processed.
- Does a built in fair housing check cover my state?
- Assume it does not until somebody proves otherwise. Federal protected classes come from the Fair Housing Act, and a built in check runs on that floor unless your state and local rules were written into it. Local class lists vary and change. A check that reports itself complete while covering only the federal list is the version to avoid, because the owner reads the line and stops looking.
- In what order should I buy tools for my business?
- Sort the whole stack before ranking anything by urgency. Sorting asks two questions of every tool: is it in the business now, and should it be. Ranking first fills the urgent list with work that should have been refused, so effort goes into improving things that should have been cancelled.
Worth passing along
This is the piece to send to somebody who has just asked what AI they should buy. It answers with layers rather than with a product name, which means it stays useful after every tool in it has been replaced. An agent, the lender she sends buyers to and the contractor she recommends will fill these seven layers with almost no overlapping products, and the layers themselves will be identical.
Written by Brett K Moore. We build the PAD System, a records structure for people, assets and deals that lives as plain text files on your own computer. Read what it is.
Read next
How do you decide whether a tool belongs in your business?
A four box audit for a software stack: what is there and should be, what is missing and should be there, what is there and should go, and what is missing for good reason. Written for real estate agents, brokers, mortgage lenders and the trades who work alongside them.
Should your CRM be where your business remembers things?
A CRM is built to move people through a pipeline and it is very good at that. It is a poor place to hold what you know about a person, and the two jobs pull against each other. Where the line sits, and the one rule that stops the two systems disagreeing. Written for real estate agents, brokers, mortgage lenders and the trades who work alongside them.
How do you know your automation is really working?
A nightly job reported success every morning for seven weeks while reading the wrong table, and half a million words sat unread on the owner's own disk. What the log said, why every line of it was true, and the four rules that came out of it. Written for real estate agents, brokers, mortgage lenders and the trades who work alongside them.