Connect a database your customers share
Connect a database whose tables hold several customers' rows — directly, or through the Gateway — and approve the tables the assistant may read.
What ran
Check the isolation yourself, on your own data
Every answer that read your database opens to show the query the AI wrote beside the one your database actually ran.
Sonela keeps each of your customers to their own rows by rewriting the query, outside the AI, before it reaches your database. The security brief explains how.
This page shows you how to watch that rewrite happen on your own data, one answer at a time, without taking our word for it. It is the record the platform shows for every database answer.
Ask something only your records know
The Ask screen answers as one of your customers and names them at the top. Ask something only your records can answer — how many orders are still open, say.
Open What ran
The first query is the AI's. It names your tables and no customer at all, because the AI is never told who is asking.
The second is what your database was given: every table replaced by one holding only that customer's rows, the customer passed as a separate value the AI never sees, and the row limit applied. Look for the customer named at the top of the screen: it appears in neither.
An example
The same question, before and after Sonela
What the AI wrote
SELECT "Status", count(*) FROM "Orders" GROUP BY "Status"
What your database was given
SELECT "Status", count(*) FROM (SELECT "Id", "PlacedOn", "Status", "Total" FROM "Orders" WHERE "TenantId" = @__ai_tenant AND "IsDeleted" = false) AS "Orders" GROUP BY "Status"
Only the columns you approved are read, deleted rows stay out, and
@__ai_tenant is filled in by Sonela from the signed-in session, never by
the AI.
In the panel
Everything a reviewer would ask about that answer
What became of each query
Whether it ran, or was turned away before it reached your database and why — a query that tries to change anything never gets that far.
How your database was held to reading
The tables it read, and how: inside a read-only transaction, rolled back, on PostgreSQL and MySQL; on SQL Server, which has no read-only transaction, checked as read-only and rolled back. Beside them, the manuals the answer was given.
A copy for your review
One button copies the whole record, so it can go straight into a security review or a ticket.
A query you can hold it to
Save as verified keeps the query with a golden question, and every later replay of that question is held to it: how it stays right.
Who sees it
You, in your dashboard, for that one answer
The queries name your tables, so they are shown to you and to no one else: the assistant your visitors use answers the same way and shows them the answer alone. The panel is made for the answer it sits under, and Sonela does not save it.
Through the Gateway, the rewrite happens on your own server. The panel then shows the query the AI wrote and says that the Gateway checked it, scoped it and ran it inside your network — the scoped query stays there.
With personal data masked, a name the AI only ever saw as a placeholder stays a placeholder in the panel, and where showing a query would show a personal value, the panel says so instead.
Don't take the brief's word for it. Ask your own database
Open a workspace, connect a database, and open What ran under the first answer that reads it.