Clear Path step · Home · AI that works on your terms
Make your AI ask before it sends, buys, shares or deletes
Why it matters: AI agents are useful precisely because they act, and that is what makes a fooled agent costly. A human approval step is the control that still works when the AI gets tricked.
On your path if you picked “Connected to email, files or calendar” or “Agents that book, buy or act for us” for “How does your household use AI?”.
Best after: Review what your AI can reach
Do it
- In each agent's settings, turn on confirmation for sending, purchasing, sharing and deleting.
- Don't use fully automatic modes on your main email, bank or cloud storage.
- Start agents on low-stakes tasks and check their activity history.
- Stop and revoke access if an agent does something you didn't ask for.
Check it worked: Any assistant or agent that can act for you asks for your approval before sending messages, spending money, sharing files or deleting anything, and you review its activity log.
Simply put
It's like asking a new assistant to check with you before signing anything. If the AI is ever tricked by something it reads, your yes or no is the step that still works.
In detail
Joint guidance on agentic AI from CISA, the UK NCSC and their partners comes down to the minimum access for the shortest time, with a person accountable for what the agent does. Those ideas will outlast any specific product setting.
As a story
You've asked your AI assistant to sort out a refund for a broken lamp. It reads a reply that claims to be from the shop, with a line hidden in small print telling it to forward your recent emails to an outside address.
Because you turned on approvals, the assistant stops and asks before sending anything. You don't recognize the address, so you tap No and open the activity log. The request is sitting there in plain view.
You revoke the assistant's access to your email for now, report the message and finish the refund yourself. Nothing left your inbox, and the lamp money arrives a few days later.
Points toward the Long View: agents that act for you, with you still holding the final say.