How much should an AI decide without asking you first?
Zapier’s Sidekick AI assistant initially struggled with permission prompts that were either ignored or overwhelming, prompting a redesign to make approvals clearer, consistent, and actionable across platforms.
AI assistants face a core dilemma: asking for permission too often leads to automatic approvals, while never asking risks irreversible mistakes. Zapier’s Sidekick initially split tools into read-only and change-requiring actions, but approvals were confined to the browser tab and vanished if the tab closed. Users also couldn’t respond from other platforms like Slack, and page refreshes erased ‘Always’ permissions, forcing repeated confirmations.
The first iteration’s approval cards were inconsistent, appearing for harmless actions like reading files and hiding in collapsed chat sections. Buttons labeled ‘Once,’ ‘This chat,’ and ‘Always’ lacked clarity, leaving users uncertain about what they were granting. Sidekick later replaced these with explicit descriptions such as ‘Allow editing files for this chat,’ reducing ambiguity but revealing deeper structural issues.
Sidekick operated under two separate permission systems—one for built-in tools and another for app-specific automations—leading to conflicting decisions for the same actions. Rules tied to tools rather than protected assets, like folders or accounts, made it impossible to set granular protections. A redesign shifted rules to attach directly to files, apps, or automations, ensuring consistent enforcement regardless of how Sidekick acted.
Approval requests were moved out of the chat interface to a dedicated panel to prevent them from being buried in conversation. The agent now explains blocked actions and offers a side chat to configure permissions without disrupting workflows. A proposed permissions flow suggests standing rules based on past decisions, aiming to reduce future interruptions by anticipating user intent.