Hi Jan,
Here’s a real example from the book. An org connected an LLM to their database and gave it column names like tbl_cust_v2 and col_rev_mtd. The agent produced queries. The queries ran. The answers were wrong.
Not because the LLM was bad. Because it had no way to know that col_rev_mtd meant “month-to-date revenue,” or that tbl_cust_v2 replaced the original table after a schema migration.
The fix is the semantic layer. Not a BI tool, a machine-readable vocabulary layer that tells AI systems what “active customer,” “churned,” and “revenue” actually mean in your business context. With it, agents stop guessing. Without it, they never stop.
|
|
|
|
|
|
