Permission check
Capability does not equal permission.
Permissions define what data, tools, systems, and actions an AI application or agent is allowed to access.
A capable AI system should still operate within clear boundaries.
Capability does not equal permission.
View information.
Example: read invoice
Change information.
Example: update CRM record
Trigger an external action.
Example: send email · release payment
Give the AI only the access required for the task.
READ invoice
READ policy
DELETE invoice
CHANGE bank account
RELEASE payment
A tool exists and can potentially be used.
Determines whether a specific action is allowed.
The email tool exists. The agent may draft an email; sending may require permission or approval.
System authorization rule.
A person reviews a selected action.
The agent may propose refunds. Refunds over $500 require manager approval.
read invoice
check ERP
flag mismatch
change bank account
release payment
delete transaction
Unless: specific permission + role + approval where required.
Safety comes from layered controls, not from trusting the model alone.
A tool is a callable capability; permission decides whether this action may use it.
Permission authorizes access; a guardrail is a broader control for inputs, outputs, or actions.
Permission is a system rule; approval is a person reviewing a selected action.
Having a tool does not permit every operation, record, amount, or recipient.
Permissions define what an AI system is allowed to access and do, even when it has the capability to do more.
A video explainer for Permissions & Safety will appear here.