This image is AI generated. In this example, AI allows people to produce photorealistic marketing visuals without a budget.
Every new AI startup, tool, wrapper, and layer is on a train. The people who laid the track have stopped laying it, and no one told the train.
Half of what has been expounded on LinkedIn in the last six months is already obsolete. I am not saying that to be mean. I am saying it because I watched it happen from my own counter.
This week I sat on a call with two other business owners. We are not programmers. We spent an hour comparing the software and tools each of us had built in the last few days.
Nothing we bought, not counting the AI subscriptions. Software we built, for our own counters and back offices, by simply saying what we wanted. Somebody will tell me that "vibe coding" is bad. Go ahead. It does not change a thing in the story.
How it actually works
Let me lay out the concept. I keep a folder. It holds how my business runs: the menus, the vendors, the decisions I made and why I made them. That folder is my context. Context might very well be one of the most valuable things we have. If tokens are the new currency, the most valuable thing to buy with tokens is context.
I open up the desktop app and start talking about the thing I want. The LLM reads the folder, writes the page, and connects it to my point of sale through the same developer tools any software company uses. I look at the result in a browser. If it is wrong, I say so in plain English. Ten minutes later it is different.
Three examples from my own list.
Restaurants sometimes have big item catalogs, mine was a monster. I go through my pos or dashboard and to make a change......I load every item, filter by category, open the item, and hunt for the field. I do the same three things to an item every single time. Maybe I have a bulk edit tool, maybe I don't. So the screen I want is one where I tap a category and see only those items and only those three fields. No other junk blocking up what I want. But remember, what I want is not the same as the hot dog guy down the hall. That is the whole application. No vendor would ever ship it, because it is shaped like me.
A shift manager at one location should not see the whole company nor need to. The old answer was a permissions table with forty checkboxes. The new answer is that I build the manager one screen that does one job. The screen is the permission. It cannot do anything I did not put in it.
When an item runs out on the line, three things should happen. It should come off the online menu. The manager should get a text. It should land on tomorrow's order list. None of that is hard. It is "if this, then that." It never got built because nobody builds for one store.
One comment on the call was that the industry is still waiting for the next programming language, the layer that sits above code and ties all of this together. I disagreed. That ship has sailed. The language is the AI itself, and we are already talking to it. Most people just have not decided what to say yet.
What changes
For twenty years, if I wanted my point of sale to do one odd thing my way, I could beg the vendor for a feature, or pay a third party for an add-on that did ninety-nine things I did not need. Now there is a third option. I build the one screen I want in an afternoon, on top of data that is mine.
What I said on that call is the whole shift in one sentence. We are not asking the vendor to build any of this. We are asking for the way to build it ourselves, on top of what the vendor already has.
That turns a product into a platform. It also puts the work where the knowledge always was. I know why my counts break. I know which three fields matter. For the first time, the person with the knowledge and the person who can build the tool are the same person.
And these tools travel. Once one owner builds something that works, it gets handed to the next owner the way a prep trick gets handed down a line. Shared, not sold.
The kitchen
I come from restaurants, so here is how it looks from a kitchen. I am not the first to borrow the kitchen for this. Technologists have written about ovens and chefs. I am writing from the other side of the pass.
Nearly every house has a stove or a microwave, and we still have restaurants.
The models are the stove. A stove gives you heat, and only a handful of companies on earth can build one. The ingredients are what you bring: your data, your customers, your context.
Knowing the recipe, heck, the hundred variations of a known recipe, is not hard. Technique spreads for free. Somebody finds a trick on Monday and every line cook has it by Friday.
What stays scarce is palate, which is knowing the dish is wrong before it leaves the pass. And a dining room, which is guests who come back.
Every restaurant person knows a brilliant cook who went broke. Skill was never the problem. The room was empty.
The train
This is bad news for a specific kind of company. The one whose whole product is "we connect to your disjointed systems and add AI." The one that says "we have the intelligence layer for your business."
The model underneath is a commodity. I can buy the same intelligence from three suppliers. If the only scarce thing you bring is the AI, your customer can now bring it too.
Ask yourself two things. Is your idea unique, or did you develop it with AI? And are you sure the AI did not do the same thing with someone else?
That is the track running out. The stove maker ships your feature in the next release. The platform you sit on top of ships it the release after that. And now your own customer can build it in an afternoon.
That last one is new. A startup used to worry about the giant above it. Now it has to worry about the owner below it, the customer it was planning to sell to. Do you know who your real competition is?
I think software is about to get restaurant economics. Low barriers, a new concept in every strip mall, high failure, thin margins, and the suppliers and landlords collecting the steady money. Every one of those new concepts is sure it is the next McDonald's.
The light
Software does not die. The add-on in the middle does. So is there a light at the end of the tunnel for a new company building on top of AI?
Yes. I can even tell you where it is.
It is in anything that needs permission. Payments, transaction records, a license, a regulated file. An agent with a browser cannot talk its way into those. The platforms that hold the payments and the records get stronger, not weaker.
It is in accountability. These products are all taking the same shape: diagnose, prescribe, act, approve. The first three are getting cheap. The seat that says "approve" or "do not approve" is the one an owner will pay for, because somebody has to be answerable.
It is in changed behavior. If what you deliver is an alert, you are finished. I can build an alert before lunch. If what you deliver is a crew that does something different on Tuesday because of the alert, that is hard, and it is still yours.
It is in palate turned into a system. A chef who cooks every plate cannot grow. The ones who last teach the line to cook the way they taste.
And it is in the dining room. Real estate and consistency built the chains. The recipe did not.
The other side
One of the owners on that call had the other side, and he was right to bring it.
His banker meets with small businesses all day. He asked him how many are having this conversation. The answer was one other. Meanwhile the biggest software companies are having their best years.
So which is it? Are owners about to build their own tools, or are a handful of us techies tinkering while everyone else keeps paying for subscriptions?
I will not tell you when. The direction looks more reliable to me than the schedule. Most owners will not build anything. They will open the box and make the model on the front, and that is fine. Somebody has to write the instructions.
The three of us are not special but we are early.
The real winners are going to be the people who can teach others how to think and how to use these new tools. Learning a programming language is over, learning how to talk out your ideas....that is gold.
The comments will be "You can't maintain that", "It costs to keep it running", "big companies don't care about that or the cost". Well, I can tell you big companies often use managed IT because they do care. Healthcare will prove that to you. They aren't buying software, they are buying someone to manage the software. How long before those managers of software realize they can just manage their own software?
I pushed this maintenance question further and asked whether, if I died, my son could keep my actual setup running with ordinary computer skills, my password book, and my phone. The AI I asked started listing objections. But here was the thing: Claude had helped build the system, and Codex had already read the repositories, traced how it worked, and understood the decisions without me explaining the implementation (my knowledge system is a topic for another day, model agnosticism is a thing). The knowledge had traveled. My son's takeover was hypothetical. One AI picking up another AI's work had already happened.
That is what gets missed in “you can't maintain it.” The AI does not disappear when the application is finished. The code is still there. The decisions, the instructions, the record of what broke and how we fixed it, all live on if the system is built right from the start. Those are things another assistant can read and work from. My son would still have to want the job, approve decisions, and put his hands on a machine when necessary. He would not have to become me first. A software company can still earn his money by delivering better results at a lower total cost. But it has to compete against that whole arrangement, including AI doing the maintenance.
If you run a business, have you built anything yourself yet?
If you sell software to people like me, what do you have that I could not make in an afternoon?
There is a light at the end of the tunnel.
I would find out what is attached to it before I walked toward it.