Every company has someone who knows.

They can tell you, without looking anything up, that an order cannot exist without a customer, that a shipping date on a cancelled order means something different, and that the number everyone quotes on Monday has been wrong since the migration.

Nobody ever asked them to write any of it down.

That is the part of your data that never got built. Storing it was solved. Moving it was solved. Processing it at a scale that would have been absurd twenty years ago was solved. None of that ever produced a plain statement of what the data means.

You can spread data across as many layers as you like. At the end of all of them, somebody still has to say what an order is, which customer it belongs to, and whether it can exist without one. That statement is the model: how you would explain the business out loud to a colleague, rather than how the rows happen to sit on a disk.

There is a lot of advice at the moment about why AI projects are failing. Your foundations are weak. Your governance is missing. Your data is not ready. Most of it is true, and none of it reaches the model itself, which is the thing all of it is standing on.

Catalogs and governance programs were built to fix exactly this, and the people who ran them were right to try. The difficulty is the shape of the work. It grows with the size of your database rather than with the questions anyone is actually asking, so it never has a natural place to stop, and a year in it is hard to say what it bought.

So we started somewhere smaller.

Point it at your database and it works out the structure: which tables relate to which, where a relationship allows only one of something, and where a value is allowed to be missing. Then it writes what it found as sentences.

Each Order is placed by exactly one Customer.
An Order can be shipped by more than one Shipper, or none at all.

Now the person who knows can read it. There is no diagram to learn and no SQL to follow, so they can simply tell you which sentences are wrong. They can change them too, and what they change is what the database is then queried by. The sentence and the rule are the same thing, so the description can never quietly stop matching the model.

We checked the assumption underneath all of this before we built on it. We renamed every table in a database to invented words, left the structure alone, and asked the same questions again. The answers still came out right. The meaning was never carried by the labels. It was in the structure the whole time, which is why the words on your model can safely be your own.

The tool is not the hero.

The person who knows is. Most of what we have built is an attempt to bring them back into a conversation they were quietly removed from, and to keep the job small enough that they can actually finish it. The facts your questions travel through, rather than an inventory of everything you own.

It shows most clearly when you have just been handed a database nobody can explain. Compile it, sit down with whoever has been there longest, and read the sentences out loud. You will both be modeling before the coffee is finished, and at the end of it you will have something you can ask questions of.

Grammar is where the model is built. It keeps checking itself against the database and shows you what moved. Vision is for seeing: the chart or the number you want in front of you right now. Reason is for the questions that start with why.

Your model is a body of declared truths. It belongs to the business, and it survives the software it was made in. Your own agent can read it over MCP, the protocol AI tools already speak, and your own software can call it directly. Either way it runs the same checks ours do.

We have worked in data and machine learning for over twenty years. The first graph compiled in March 2026, before any of this was a category. We are here for the same reason we started: this part was missing, and it belongs nearer to the people who know than it has ever been allowed to sit.