AI: Three Levels of Judgement

I have never vibe-coded. No judgement on the people who do, but I can’t put my name on something I don’t understand well enough to test. That’s the floor.

I have never agentically engineered either. This is judgement included in a judgement loop, which sounds like a hedge until you think about what it means: it’s only as effective as the judgement you brought to the table and the training of the model. You’re betting the work on variables you don’t control.

I have only ever done directed engineering. Judgement constantly present. Me, making decisions about what the system does and why. AI amplifies my skills on work I fully own.

Here’s why that matters for the kind of work I build.

Real data architecture isn’t abstract. It’s dimensional models, business vault patterns, star schemas, rules engines built on specific business logic. A semantic model is full of business encodings. The actual logic that makes a thing work lives in the specifics: how this company measures quality, how they define a closed order, what sequence of events triggers a recount, which table joins on what field and whether a left join or an inner join changes what the answer means.

A frontier AI model can wax poetic about data architecture theory. The data warehouse literature. Best practices. But ask it to reason about an entire star schema without the context, to understand business rules without you explaining them, to build metrics that carry the weight of real decisions. It can’t. Not because it’s not smart. Because the model was never trained on your business. The encoding lives in your history, your operations, your people.

That’s why every dimensional model I’ve built, every business vault layer, every rules engine started with me understanding the business logic first. Deeply. Then the work wasn’t “make AI build a data model.” It was “I know what this model needs to do, now use your speed to help me build it.”

When you hand the architecture to a model and ask it to design by itself, you get something that sounds reasonable until it runs into the first business rule nobody wrote down. Then you’ve got something with your name on it that you don’t fully understand. That’s the spot I won’t occupy.

This is why OIM is built the way it is. The operations intelligence model earns the right to be called infrastructure only because every layer of it carries business logic that was understood before it was built. Directed engineering all the way through. That’s not a limitation. That’s the whole point.