The easiest way to make an AI database agent dangerous is to let tenant scope become a suggestion.
A human analyst usually knows that a customer support question should only touch one account.
A model does not know that unless the system makes the boundary explicit.
And if the boundary lives only in a prompt, it is not a boundary.
It is a preference.
Why this matters
Most SaaS databases contain data from many customers in the same logical system.
Application code normally adds the current tenant, workspace, account, or organization filter automatically.
Natural-language SQL changes the path.
The user asks:
show me recent failed syncs
or:
which invoices are overdue?
The agent turns that into a query.
If the system does not enforce tenant scope outside the model, the agent may generate a valid query that answers the wrong audience.
The failure may not look like a crash.
It may look like a plausible answer with other customers' data included.
What good tenant scoping looks like
For AI database…