Safety and Approvals

How Blue confirms bulk and destructive AI actions before they run.

The AI Agent can do real work in Blue: create records, update fields, add comments, manage checklists, and move records to trash. Safety is handled in two layers: the assistant only uses your existing Blue permissions, and higher-risk actions ask for approval before they run.

Permissions always come first

The assistant acts as the person who invoked it. If you cannot update a record, move it, or view a workspace yourself, the AI Agent cannot bypass that restriction for you.

This applies everywhere the assistant is used:

  • The AI Agent panel.
  • @Blue mentions in record comments.
  • Actions the assistant performs after you approve them.

When Blue posts a comment or appears in activity, the action is still attributed back to the person who asked for it.

Autonomy modes in the AI Agent panel

In the panel, the autonomy mode controls how much approval is required:

  • Plan — read-only. The assistant can investigate and recommend steps, but it does not change data.
  • Ask — every write action waits for you to approve it from an action card.
  • Auto — lower-risk writes can run automatically, but bulk, destructive, or uncertain actions still require approval.

You can change the mode from the composer before sending your next message. See Using the Assistant for the full workflow.

Approvals in record comments

When you mention @Blue in a record comment, the assistant is optimized for quick work inside that thread. Simple, low-risk actions can run immediately, for example creating a record, adding a checklist item, or updating a single record.

Higher-risk comment actions create a confirmation message instead of running immediately:

  • Bulk updates across multiple records.
  • Bulk custom field changes.
  • Moving records to trash.

For example, if you ask Blue to move five records to trash, Blue replies with a summary like:

I can move 5 records to trash.
Reply @Blue approve to run it, or @Blue cancel to skip.

Reply in the same thread with @Blue approve to run the pending action, or @Blue cancel to reject it. Natural replies such as @Blue yes, go ahead also work when the intent is clear.

Ambiguous replies do not run actions

If your reply changes the scope, adds a condition, or is unclear, Blue asks for clarification instead of executing the pending action.

Examples that do not approve an action:

  • @Blue only the urgent ones
  • @Blue wait, which records?
  • @Blue maybe later

This is intentional: approvals must clearly refer to the pending action.

Who can approve a pending comment action

By default, the person who requested the action must approve it. Organization owners and admins can also approve pending actions, so they can unblock work when needed.

If someone else replies with approval, Blue does not execute the action and explains that only the requester or an admin can approve it.

Trash, not permanent deletion

When the assistant deletes a record, it moves the record to Trash. It does not permanently erase the record from Blue as part of the assistant action.

Trash still counts as a destructive action because it can affect workflows, visibility, and activity history. That is why Blue always asks before moving records to trash from comments or the panel.

Multiple pending requests

If one request creates several pending tool calls, Blue treats them as one approval group. For example, moving five records to trash produces one approval summary and one approval reply runs the whole group.

If there are pending requests from multiple separate turns in the same comment thread, Blue will not guess which one you mean. In that case, use the AI Agent panel or ask again with a narrower request.