Hex's AI agent needs an admin to mark data as sensitive

The setup

Hex is a collaborative notebook built for data analysts, and its AI product page, as retrieved 16 September 2026, describes an assistant built into that notebook rather than a separate chat window, one that understands a user's data and intent and can write and edit code while tracking a project's upstream and downstream dependencies. Using it requires connections already set up in a workspace, a warehouse connection and optionally the semantic layer Hex calls Context Studio, before the assistant has anything to reason about. A separate documentation page covers a second setup step: marking specific data connections as Sensitive, a workspace-admin-only setting.

What the documents show

The product page states the assistant can leverage context from semantic models, database metadata, external repositories, and existing dashboards when generating a query, and that all metadata is stored in a secure database inside Hex's own architecture, with Hex's model providers not training on customer data. Those are Hex's own claims, not an independent audit. The data-connection documentation adds a detail the marketing page omits: outside the main app, in a Slack thread or an MCP client session, the Hex Agent automatically selects from connections the user has access to, which the documentation itself flags as a risk, since a casual question could route analysis against data a person would rather keep off those surfaces.

The friction

That is the friction the vendor discloses directly: by default, a person's ordinary data access extends into an AI conversation started elsewhere, and stopping that requires an administrator to act, not a default-safe setting. The documentation states plainly that users require an Admin role to access that setting, so an individual analyst cannot unilaterally protect a connection they consider sensitive. Neither page states whether the assistant can write back to a connected warehouse, only that it can read metadata and generate queries, leaving write access an open question to confirm against actual permissions.

What changed in the work

Before this feature, turning a plain-language question into a SQL or Python cell in a shared notebook meant an analyst writing it directly, checking a schema by hand or asking someone who knew it. The documented capability moves that first draft to the assistant, grounded in metadata Hex says it already indexes. What the sources do not establish is that this draft is safe to run untouched: the sensitive-connections documentation exists precisely because a default, permissive routing of context was judged risky enough to need an override, which is the vendor's own evidence that the convenience and the exposure arrived together.

• Have any of this workspace's data connections actually been marked Sensitive, or does the default apply everywhere?
• Does the Hex Agent in Slack or MCP have access to connections this specific user would not normally query directly?
• What does context from external repositories include for this workspace, and who configured that scope?

A convenience feature that ships with its own admin-only override is telling you something about how it behaves before anyone flips that switch.