Most conversations about enterprise AI focus on models.
How smart they are.
How many tokens they support.
How well they generate SQL.
After working with enterprise data, I think we're paying attention to the wrong problem.
The Problem Isn't Finding Data
Enterprise AI usually has access to plenty of data.
Schemas.
Data catalogs.
Documentation.
Historical SQL.
Yet it still struggles with surprisingly simple business questions.
Why?
Because it doesn't understand how the data is connected.
A Simple Example
Imagine asking an AI agent:
"Show the top customers by revenue."
It scans the database and finds:
- 4 customer tables
- 3 revenue-related tables
- multiple possible join keys
From the model's perspective, several SQL queries look perfectly valid.
Only one matches how the business actually defines revenue and customers.
The model can't infer that from table names alone.
Relationships Carry Business Knowledge
This is something I've started appreciating more over the past year.
The relationship between tables isn't just a technical detail.
It's business knowledge.
A trusted relationship tells you:
- which table is the source of truth
- which join path has been validated
- which fields represent the same business entity
- which datasets should never be joined together
Without that context, AI is forced to guess.
Sometimes it guesses correctly.
Sometimes it doesn't.
Bigger Models Don't Fix Missing Context
Every new model is better at reasoning.
That's great.
But reasoning only works when the underlying context is reliable.
If the relationships are ambiguous, a more capable model simply produces a more convincing wrong answer.
That's why many enterprise AI projects spend far more time validating results than generating them.
Final Thought
I'm becoming convinced that enterprise AI isn't just a language problem.
It's a data relationship problem.
Models will continue to improve.
The bigger opportunity is helping them understand how enterprise data actually fits together.
Because once AI understands relationships, everything else becomes much easier.
Top comments (0)