If your semantic layer disappeared tomorrow, could you rebuild it?

AI has made writing SQL much easier.

But writing SQL was never the hardest part of analytics.

The harder part is knowing what the business actually means by things like ARR, churn, active customer, or adoption, and making sure everyone calculates them the same way. That’s why semantic layers have become so important.

I recently came across an interesting experiment from MotherDuck where AI agents were used to generate a semantic layer from business context and a set of questions with expected answers.

What I found more interesting wasn’t whether AI can replace the semantic layer.

It was this: if implementation becomes easier to generate, what should data teams actually preserve?

Maybe the most valuable asset isn’t thousands of lines of SQL or YAML.

Maybe it’s the business context, definitions, decisions and tests that allow those implementations to be rebuilt reliably.

A useful question for data teams:

If your semantic layer disappeared tomorrow, would you have enough documented business knowledge to recreate it?