As engineers, we spend a lot of time talking about AI models.
Which model generates better SQL?
Which model reasons better?
Which one has the largest context window?
But after working with enterprise data, I've started to think we're looking in the wrong place.
Most enterprise databases were never designed for AI.
They were designed for applications.
## Applications Know the Rules. AI Doesn't.
A business application already knows where everything is.
If an order needs a customer record, the developer has already defined the relationship.
If a dashboard needs revenue, someone has already decided which calculation to use.
The application doesn't need to discover anything.
AI does.
When an LLM connects to an enterprise database, all it sees is hundreds of tables and thousands of columns.
It has no idea:
- Which customer table is authoritative.
- Which tables are safe to join.
- Whether two IDs represent the same business entity.
- Which revenue definition the business actually uses.
Generating SQL isn't the difficult part anymore.
Choosing the right data is.
*## Schemas Describe Structure, Not Business Knowledge
*
Even well-designed databases have this problem.
A schema tells you that a table exists.
It doesn't tell you:
- why it exists,
- when it should be used,
- or whether another table has replaced it over time.
The knowledge that engineers build up over years of maintaining a system rarely exists inside the database itself.
It's stored in documentation, meeting notes, old dashboards—or simply in someone's head.
That's exactly the information AI is missing.
## Two Things Make Enterprise Data More Understandable
In my experience, AI becomes much more reliable when two gaps are addressed.
First, data relationships.
AI needs to know how tables, fields, and business entities are connected—not just through foreign keys, but through relationships that have been verified across real enterprise systems. Discovering and validating those relationships is the foundation of platforms like Arisyn-IntaLink. :contentReference[oaicite:0]{index=0}
Second, business semantics.
Even after the right data is found, AI still needs to understand what that data means. Shared metric definitions, business terminology, and governed semantic rules help ensure that "Revenue" or "Customer" means the same thing to everyone. That's exactly the role of a semantic layer such as Arisyn-Semora. :contentReference[oaicite:1]{index=1}
Relationships explain how data is connected.
Semantics explain what the data means.
AI needs both.
## Final Thoughts
I don't think enterprise AI is limited by SQL generation anymore.
The bigger challenge is helping AI understand enterprise data the way experienced engineers do.
The better we capture relationships and business semantics, the less AI has to guess.
And in enterprise systems, fewer guesses almost always lead to better decisions.
Top comments (0)