Agent Rules · Version 1.0 · November 2026
Eight rules for AI agents that work inside a business.
AI agents can now answer customers, update websites, and change bookings. These rules are the standard I design client systems to, and you are welcome to use them with anyone. A link back is appreciated.
- 1
Every agent has a named owner
- Why
- When an agent makes a mistake, someone must notice, decide, and fix it. "The AI did it" is not an answer a customer accepts.
- How it is built
- Each agent is assigned to a person on your team who receives its approval requests and its monthly summary.
- 2
Agents draft; people approve what matters
- Why
- Anything a customer sees, anything that costs money, and anything hard to reverse carries your reputation.
- How it is built
- Customer-facing messages, published content, prices, and bookings start as previews. A person approves them. Categories move to automatic only when the owner decides they are safe.
- 3
Least access, by default
- Why
- An agent that can do everything will eventually do something it should not.
- How it is built
- Agents get only the tools their job needs, with separate read and change permissions, and credentials that can be revoked without affecting anyone else.
- 4
Every action is logged in plain language
- Why
- Trust comes from being able to see what happened, not from being told it is fine.
- How it is built
- Each request records what the agent did, who asked, when, and the result, in words an owner can read without a developer.
- 5
Everything can be undone
- Why
- Mistakes are inevitable. Damage is optional.
- How it is built
- Changes go through previews and versions, retried requests cannot run twice, and every change has a way back.
- 6
Agents answer from your source of truth
- Why
- A confident wrong answer about a price or a schedule is worse than no answer.
- How it is built
- Agents read from your own website, CMS, booking system, and documents. When the answer is not there, they hand over to a person instead of guessing.
- 7
Measured every month
- Why
- An agent that is not reviewed slowly drifts away from how the business actually works.
- How it is built
- A monthly review covers what agents handled, what people changed or rejected, any incidents, and what to adjust next.
- 8
Anyone can press pause
- Why
- The person closest to the customer often sees a problem first.
- How it is built
- Any member of staff can pause an agent immediately, without technical help, and the owner is told why.
Put them to work
Want agents that follow these rules in your business?
The AI Operations Audit includes a permission policy built on these rules for your team. Questions or suggestions for the next version: email me.