Human in the Loop
Perplexity surfaces Gmail and Outlook connector chips in the composer and lets users Allow or Disable individual read and write tools.

Gmail and Outlook connector chips appear when query implies email context
What's happening
Perplexity front-loads human-in-the-loop consent via connector chips and granular tool permissions before agents access personal data.
Patterns
Read-only vs write/delete tool groups with per-capability Allow controls
Users opt into Gmail/Outlook before Computer mode acts on mail
UX Insights
- •Connector chips appear in context when queries touch personal data
- •Tool permissions split read vs write risk tiers
- •Users can Allow capabilities individually instead of all-or-nothing
Design Decisions
Perplexity treats agent access as a configurable permission matrix, not a single OAuth checkbox.
More case studies

Human in the Loop
ChatGPT clarifies before acting: structured questions for bookings, email draft cards with edit/export handoffs, and explicit limits when it cannot complete an action.

Human in the Loop
Gemini asks clarifying questions before drafting email, previews Gmail drafts in a card, and gates Send behind explicit user approval.