AI2SQL masks results before they reach the assistant, but that does not mean data stays inside the database
AI2SQL performs MCP result masking on its servers, allowing an assistant to receive placeholders instead of original customer values. This suits analysis that still needs real data for queries. But it does not replace database permissions or mean raw data never reaches AI2SQL; using this capability with your own database requires Team.
If a team wants an AI assistant to analyze orders without directly seeing customer email addresses and phone numbers, AI2SQL’s result masking is worth evaluating. The key is understanding where it takes effect: queries still use real data, and results are processed on AI2SQL’s servers before being passed to the assistant. If an organization requires that even the AI2SQL service cannot access original customer values, this capability alone does not provide the required data boundary.
This feature analysis is based on the official AI2SQL MCP documentation; it has not been tested with real customer data. According to the vendor, after masking, the assistant can receive placeholders, partially masked values, or fields with their values hidden, while original values are no longer passed to it. That commitment concerns delivery of results and should not be interpreted more broadly as meaning the data never leaves the source database.
Query real data and deliver processed results
The official documentation states that queries run on real data, so counts, joins, and filters remain effective; masking occurs before the returned rows are passed to the assistant. This allows “count orders by region” and “list customer contact information for the assistant” to use different rules for delivering results.
For example, if an operations employee only wants to know which regions have the most pending orders, there is no need to include customer names and email addresses in query results. First restrict the query itself to the fields required, then use masking as an additional safeguard. If the task requires checking a specific record, assess whether placeholders or partially masked values are sufficient to do the work, rather than providing full contact details to the assistant by default. These are suggested analysis scenarios, not hands-on product test cases.
The granularity of settings determines what the assistant can ultimately see
The MCP page lists rules including placeholders, partial masking, hiding, and showing. These can be set by column, with a preview available before saving. Masking can also be enabled at the connection level. For a team, “has a masking feature” is only a starting point: inspect the actual rules applied to every connection and field you intend to make available.
Build an acceptance-test sample using synthetic data: a regular email column, contact information in notes, and category fields that must remain available for analysis. Check that the preview matches the results the assistant actually receives, and confirm that a “show” rule does not release values you intended to hide. The vendor also describes detection of sensitive patterns in free text, but that description cannot replace validation with the team’s own fields and data formats.
When connecting your own database, result masking is a Team feature; the official sample database, demo_shop, allows it to be tried on every plan. Seeing masked results in the sample does not prove that your current plan covers your production connections. Confirm both the required plan and the settings of the actual connections during purchasing.
Read-only access, masking, and access scope each serve a different purpose
The same page states that the query service rejects writes, DDL, and multi-statement batches on the server side. Read-only restrictions address data modification, and result masking addresses the values the assistant sees. Neither should replace least-privilege permissions at the database level. Teams should still make available only the tables, views, and fields required for the task, and verify that the service’s processing of this data meets their requirements.
This capability suits teams that allow AI2SQL to process query results but want to reduce downstream assistants’ access to original customer values. If most work only requires aggregate metrics, returning fewer fields is usually more direct. If the business needs to inspect complete identity information record by record, masked results may not support the task. Distinguish these two needs when choosing the capability so you can tell whether it reduces unnecessary exposure or also hides information needed for the work.
Content note:This article is based on publicly available information and is intended as a reference for evaluating tools. Features, pricing, and related policies may change at any time; please refer to the latest official information.
Original source:ai2sql.io