The AI is the easy part

Here is the thing almost everyone has backwards right now.

The AI is the easy part. It is a commodity. I can rent the same models you can, from the same handful of companies, for the same price.

Anthropic and OpenAI have done something genuinely remarkable, and none of it is a moat for the people building on top of it. If your plan is that the model is your advantage, you do not have an advantage.

You have a subscription.

The hard part is the part nobody wants to do. It is the structure that tells the model what a performance problem actually looks like inside your business.

Skip that, and a very powerful model will confidently hand you very generic answers. Most companies are pouring their effort into the half that is already solved and skipping the half that is the whole game.

This was already broken before AI showed up

Take AI out of it for a second. Pretend it does not exist. You still have the problem.

For twenty years the pitch for dashboards and BI was simple: put the numbers in front of enough people and they will make better decisions. If everyone is honest, that mostly did not happen.

Dashboards have to be generic, so they point at obvious, high-level stuff that nobody can act on. And two gaps sit underneath the whole thing.

• The technical people know how to drive the tools, but they are not steeped in how the business actually runs.
• The operators know exactly how the business runs, but they are not technical, and the tools were never built for them.

So what does everyone do about it?

They try to train people to survive Power BI. And since nobody is sure what any given person needs, they load them up with 150 reports and put 50 prompts on each one, just in case someone asks the question.

The people were never going to ask the question. That is the tribal-knowledge problem in one sentence: the judgment that actually runs the business lives in a few people's heads, the organization has no structure for writing it down, and every new hire takes six to twelve months to absorb even a piece of it.

That judgment is real.

The operators who have been there fifteen or twenty years know the patterns. They know what to look at first, what a dip means here versus there, which problems are worth chasing and which ones fix themselves. Nobody ever wrote it down because there was never a way to.

What generic AI does when you skip the structure

Now put AI back in. This is exactly the kind of reasoning and judgment work AI is supposed to be good at.

But generic AI cannot do it on its own, and the reason is not a knock on the models.

Pre-training is built on what sits on the public internet. The knowledge that runs your business is not on the internet. It is not written down anywhere a model could scan. When you get hired into an industry, you do not learn it from a website. You go work there for years and put it in your head.

That is exactly why people who move company to company inside an industry are so valuable, and it is exactly what a general model has no access to.

So here is what most people build when they try to close that gap themselves. They have no strong opinion about the specific problem, so they create a giant dumping ground.

Throw everything into a vector database and a RAG index and hope the model sorts it out.

That's better than nothing. But that's not structured context.

It is a search box. Sometimes you get lucky. But the model still does not understand the domain, it still makes things up, and it tries a different approach every time you run it.

In operations that is disqualifying. If a signal means trouble in one location, it has to mean trouble in the next one, or nobody trusts any of it.

Anyone in analytics will tell you governance and trust are most of the job, and you do not get either one from a pile of documents and a good model.

A data model, but for judgment

Here is the reframe, and I want to state it plainly instead of making you dig for it.

In the older world of software, every serious application has a data model underneath it. Salesforce has one. ServiceNow has one. SAP has one.

How Salesforce decided to structure what an opportunity is, versus an account, versus a contact, is not a footnote. It is a huge chunk of the intellectual property.

The application is valuable precisely because somebody made those structural decisions once, so you are not rebuilding CRM from scratch every time you log a deal.

An AI application needs the same thing, one level up. Not a data model. A context model.

A structured framework for how you find performance problems, how a diagnosis turns into a recommended action, and how that logic applies at every level of an organization.

It is designed for one job, driving performance management from data, and it holds an opinion about that job. It is not a warehouse for every file the company owns.

The discipline is to start from the finish line and work backward. You do not start with every document and hope it solves everything. That gets you nowhere.

You start from the goal, finding and fixing performance opportunities, and you build only the structure that serves it. The overwhelming majority of a company's documents are irrelevant to that goal.

Leaving them out is the point, not a limitation.

The architecture is the moat, not the model

This is the part that took four years and does not come out of a prompt.

Capturing the context is not a survey. We sit with the operators and literally record them going through their own numbers, explaining how they read them, where they look first, what patterns they chase.

Different leaders hold different opinions, so we consolidate all of it into one structured model rather than a stack of contradictions.

Then we show it back on real locations they know cold, because people only react once they can see the thing written down. Roughly the bulk of it lands as "yes, that is what I would have said."

The rest is exceptions and edge cases, and those get built into the framework. Then it runs again.

That loop is the work.

Underneath, the context feeds an engine that decides which investigation paths to take, executes the actual analysis, and puts every output through automated trust checks before a single manager reads it.

The layers matter, but the point is simpler than the diagram: it is a lot of deliberate structure, built and refined over years, aimed at one problem.

You cannot prompt your way to that. You cannot dump your way to it either.

It is an accumulation of specific decisions about a specific problem, which is exactly what an application is and exactly what a pile of documents is not.

Why this is the whole argument

Once the context model exists, it becomes a harness for running the same judgment consistently, across every location, every cycle, in a way you can govern and trust.

When something means trouble in one place, it means trouble everywhere.

That is what turns a clever one-off analysis into a system a COO can actually run a distributed business on. And it compounds.

The things that do not change get codified once and stop eating anyone's attention, so your people spend their time on what is actually different this period.

So when people tell me it is all the same thing now, that AI is AI, I think they are missing the only distinction that matters.

I think of generic AI like a database. It's a tool. But a database doesn't encapsulate workflow and processes and businesses. That's why you have an application. It's the difference between Oracle and Salesforce. We're the Salesforce layer. Claude and ChatGPT are the database.

The database is a commodity. The application is the moat. Almost everyone is building the wrong half.