Back to Glossary
AI Permission Design
Synonyms: agent permissions UX, AI consent design, human-in-the-loop approvals, agent authorization UX, tool access controls
Do not index
Definition
AI permission design decides what an AI agent can access, what it can do with that access, and when it has to stop and ask the user first. It covers the full cycle: the access request, the scope of each grant (read vs. send, one folder vs. the whole drive), mid-task approvals, and the page where people revoke access later. It exists because agents act with the user's real accounts.
Use cases
Ask too little and the agent does things no one approved. Ask too often and people approve everything without reading.
- The all-in-one connector: An agent meant to summarize email can also send it. One hidden instruction in a spam email, and it forwards private threads to a stranger.
- Approval fatigue: A coding agent asks before every edit. By prompt 30, the developer approves a folder deletion without reading.
- The forgotten grant: Folder access given for one task is still live months later, and no screen lists it.
How it's used in practice
- Ask at the moment of need: Ask when the task starts, with one line on why.
- Split read from write: "Read my calendar" can run freely while "send an invite" still asks.
- Give each tool 3 settings: Always allow, Ask first, Block.
- Approve the plan once: Show multi-step plans up front, get one approval, keep a stop button.
- Ship a permissions page: Every grant, its last use, and a one-click revoke.
Tier approvals by how hard the action is to undo. Reads run freely. Sends, payments, and deletes always ask and name the target: "Send this draft to jane@acme.com?" Enforce the tiers on the server too. A UI-only confirm button stops nobody.
Challenges & limitations
- Control costs speed: Most commercial agents offer constant approvals or AI auto-reviewers with little user control (UW, 2026).
- Actions chain: One approved step can trigger others the user never saw.
- Prompt injection still works: Hidden instructions can steer an agent within its access. Tight scopes cap the damage.
Free resources
- NN/g: 3 Design Considerations for Effective Mobile-App Permission Requests — copy, timing, and decision reversal for permission prompts, the groundwork agent prompts build on.
- Smashing Magazine: Designing For Agentic AI: Practical UX Patterns For Control, Consent, And Accountability — Victor Yocco's 6 agent patterns, including Intent Preview and Autonomy Dial, each with success metrics.
- Anthropic: Trustworthy agents in practice — how per-action settings (always allow, needs approval, block) and Plan Mode keep users in control of Claude.
