I have been a real estate investor for over twenty years. The first home I bought was in Plymouth, Michigan. Since then, I have been an investor in multi-family. I started building with AI when ChatGPT shipped in November 2022. The first practice is more than eight times older than the second. The first practice is teaching the second most of what I know.
When I sit down with the model to write a spec, the spec is in the language of someone who has been a real estate investor for twenty years. When I make a decision about whether to use a particular AI tool in my work, the decision runs through the same screen I use when I look at a property. The homes taught me how to read for what is real and what is staged. The homes taught me to ignore the brochure and ask about the roof. The homes taught me to look at the basement.
This essay is about what twenty years of being a real estate investor has been teaching me about being an AI architect. The two practices are not as separate as the categories make them sound. They are both about owning something for the long hold and watching it work, over time.
What the first home taught me, Plymouth, 2008#
I bought my first home in 2008, in Plymouth, Michigan. A four-bedroom on a cul-de-sac, in a good school district, on a street with trees that have taken fifty years to grow. Downtown amenities were a short walk away. I did not know what I did not know.
What I learned in the first year was that the home was not what I had bought. I had bought a relationship with the people who lived in it. I had bought a relationship with the neighborhood. I had bought a relationship with weather, with plumbing, with the way the property tax bill arrives in January and demands attention by February. The home was the asset on the spreadsheet. The home was also a family, a school, a street. The relationships were the work.
The second thing I learned was patience. Real estate numbers do not move in a quarter. They move in a decade. The rent goes up by a small percentage. The value of the building goes up by a small percentage. The mortgage gets paid down by a small percentage every month. Compounding is the only force that makes any of it work. Compounding does not work in a quarter.
The third thing I learned was the difference between the building you maintain and the building you neglect. The difference is invisible in year one. The difference is the property in year fifteen. You do not see compounding while it happens. You see compounding when you look at the photograph of the property in 2008 next to the property today.
The first home led to another. The second home led to a portfolio. I moved from single homes into multi-family because the operations could be run together. The discipline got bigger. The lesson did not change.
I have been making these three lessons for twenty years. Each property added to the lesson. Each tenant added to the lesson. Each repair I almost deferred and then did taught me something I would have been embarrassed to need.
The lessons did not transfer to AI at first. They transferred when I noticed I had been applying them without naming them.
The ChatGPT moment, the scope of work#
I sat down with ChatGPT in November 2022. I asked it a question. The answer surprised me. I asked it another question. The answer surprised me again. I have built systems that produce these answers. The system on the other side of the chat window was different from anything I had built before. I knew that everything I was doing in my technology work was about to change.
The first weeks were chaos. I tried to use the model for everything. Most of it did not work. Some of it worked too well and got me into trouble. I was treating the model the way a first-time investor treats a new home. I was getting too involved with the details and missing what the property was for.
The shift happened in the third month. I was about to draft a long memo with the model. I stopped before I sent the prompt. I closed the laptop. I went and got a notebook. I wrote the spec for what I wanted the memo to do, the way I would have written a scope of work for a renovation. Then I gave the spec to the model. The memo the model returned was the memo I had been trying to draft for two weeks. The spec was the work I had been missing.
What changed for me was the moment I started treating the model the way I treat a building. I stopped asking it for help in real time. I started writing the spec for what I needed. I started writing the audit criteria. I started giving it the same long-form work I would give a contractor working on a property. A clear scope. An inspection at the end. A punch list. A release of payment when the work passed.
That is the discipline I wrote about in the spec-and-audit essay. I did not invent it. I had been using it on buildings for two decades. The buildings had taught me what to do with the model before the model existed.
The lesson under the lesson is that the AI discipline is not a new discipline. It is the old discipline applied to a new asset class.
What the two practices have in common, long hold, specification, maintenance#
A home and a model are the same asset in three ways.
Both reward long-term thinking. The right question to ask about either is not what it can do this quarter. The right question is what it will be in ten years if I take care of it.
Both require specification. A home has plans. A renovation has a scope of work. An AI workflow has a spec. The discipline of specification is the difference between something you own and something that owns you. The spec is what makes the asset usable to someone besides yourself.
Both require maintenance. A home you do not operate becomes a liability faster than a home you do not own. A model workflow you do not audit becomes a liability faster than a workflow you do not have. The maintenance is invisible until it has been skipped for too long. Then the maintenance is the only thing anyone can see.
The homes taught me to evaluate AI tools by looking at them the way I look at properties I am about to invest in. What is this asset for. Who depends on it. What does it cost to keep it running well. What will it cost to neglect it. What is the exit. The questions are the same questions. The questions work on buildings. The questions work on AI tools. The questions work on most other things people think of as investments.
Running a property is a daily discipline. Running an AI workflow is a daily discipline. The day-to-day discipline of running an asset is the same whether the asset is a four-bedroom on a cul-de-sac or a model workflow on a laptop. The investor runs the asset. The asset does not run the investor.
Most of what I know about running the recursive audit on a model output came from running the same audit on a roof I was about to replace. The roof did not need a model. The model needs the discipline that the roof taught me.
What the homes teach that the AI rooms do not, the boiler does not read TechCrunch#
The AI rooms are loud. The homes are quiet. That is the first difference.
The AI rooms talk in quarters. The homes talk in decades. The homes do not care about the next funding round. The homes care about the boiler that is fourteen years old. The boiler will need to be replaced in year sixteen. The boiler will need to be replaced whether or not there is a new model release that quarter. The boiler does not read TechCrunch.
The AI rooms talk in users. The homes talk in tenants. The homes talk in the families inside the tenants and the schools nearby and the streets the families live on. There is a difference between a user and a tenant. A user is a metric. A tenant is a person who lives somewhere, who pays you on the first of the month because of the trust between you, who calls you when the pipe freezes, who has a child in a school nearby, who is part of a neighborhood that you are part of by extension. A user can churn. A tenant has a lease. The lease is not paperwork alone. The lease is the structure that lets the relationship outlast a bad week.
The AI rooms talk in scale. The homes talk in community. I have invested in homes in neighborhoods where I knew the names of the people on the block within six months. The neighborhood improved over twenty years because the neighbors all stayed and put in their work and watched out for each other. No model can run that pattern at scale. The pattern works because of the patience and the rootedness, not in spite of them.
The thing the homes taught me that the AI rooms are missing the most is that the people who own the asset have to know the people who use it. The relationship is the asset. The asset is not separable from the relationship. The model that gets that wrong is the model that gets shipped and then abandoned. The home that gets that wrong is the home that gets sold at a loss in year seven.
The lesson for AI from an investor’s seat, build for the long hold#
If I could tell the people building AI one thing I have learned in twenty years of investing in homes, it would be this.
Build for the long hold. Do not build for the quarter. Do not build for the demo. Build the thing you will still be willing to own in fifteen years, including the parts that are inconvenient now and will be more inconvenient then.
Audit your assets. The model you ship today is the model you will be supporting in five years. The customer who uses the model today is the tenant who is depending on you in five years. Do the audit before you ship. Do it again every year. Replace the parts that are showing wear. Do not let a small problem become a structural one.
Know the people who depend on what you build. Most AI tools are being shipped by people who have never met the customer in person. Most investors who hold properties for the long term have met every tenant in person. The meeting changes what you ship. The meeting changes what you do when something goes wrong. The meeting is the part of the work that the spreadsheet does not capture.
The AI conversation right now is full of people who have never put a roof on a home. The conversation needs more people who have. Not because the homes are sacred. Because the homes teach the long-form responsibility that the models still cannot reach.
This morning I drove past one of my homes. The lights were on in the kitchen. A tenant was making breakfast. The neighborhood was quiet. The boiler is fourteen years old. The roof is good for another nine. The spec for the next AI workflow is open on my desk when I get home.
Both assets are mine. Both assets are working. Both require the same hands.
Both.
