How the trust score works
Every use case gets a score from 0 to 100. It does not measure how smart the agent is. It measures how much damage it could do on a bad day, and how easy it is for you to catch and fix.
Access (up to 30 points)
Read only scores highest. Writing drafts or records scores less. Anything that can send messages or spend money scores least, even when a person approves it.
Human approval (up to 30 points)
Actions you cannot take back (send, pay, delete) wait for a yes. The confirm step says in plain words what cannot be undone.
Undo and logs (up to 20 points)
You can see everything the agent did and roll it back. Drafts you can throw away score higher than changes to live records.
Scope (up to 20 points)
The agent has its own login with only the permissions it needs, instead of borrowing your full account. Fewer connected tools scores higher.
Who scored these
The use cases on this site are setups we wrote ourselves. Each one is labeled Reviewed by Trustcase (example setup) and is scored only from what the entry describes. They are not reviews of any vendor's product.
Builder submissions
When builders start sending in their own agents, each score will carry one of two tags. Self-reported means the builder filled in the rubric. Verified means we checked the setup ourselves.
How to raise your score
- Access: let the agent draft, and let a person press send or pay.
- Human approval: put a named yes in front of every step you cannot take back, and say so plainly on the confirm.
- Undo and logs: have it propose changes instead of editing live records, and log everything with a way back.
- Scope: connect fewer tools, each with its own limited login.
Every use case page lists the exact fixes for that setup and the score they would reach. Want a hand? Get an AI Trust Audit from Pixeldesigns.
Reading the colors
Green is 85 and up. Amber is 75 to 84. Red is below 75.