Five controls to set before you let an AI agent buy, pay, or cancel anything for you, so it finishes the job without spending money you didn't approve.
Your agent can plan the perfect trip. -- but whether it gets to book the hotel is up to the hotel, and after Black October, the hotel has questions.
TL;DR
- AI is making it easier to break into websites, raising concerns about security when delegating tasks to AI agents.
- Businesses need methods to allow AI agents to perform actions like booking flights and managing accounts without compromising security.
- The internet is expected to become more uneven for AI agents through 2027, with varying levels of business acceptance and access.
- Companies that make legitimate agent activity recognizable and acceptable will gain power.
- A practical consequence is AI agents being unable to complete bookings on certain business websites.
- The article will detail trust systems like Cloudflare's, Google's AP2, Stripe's payment tokens, Visa's pilot, and Meta's Sentinel.
- It will explain why agents might stop working, categorizing issues into three zones for correction, better information, approval, or alternative routes.
- The author will share how they split work across five agents: Grok Bot, dots, Instinct, Muse, and Duet.
- Five controls for a first delegated job will be presented, including narrow access, specific approvals, and result checking.
- A prompt to pick an agent will be provided, designed for a small, verifiable first experiment.