A prompt can describe who may cancel an appointment, but only the execution service can enforce that rule. Keep tool permissions in a backend policy boundary that the generated conversation cannot rewrite.
Permission checks answer a different question
A tool permission check determines whether the current principal may perform a specific action on a specific resource under current conditions. Prompt instructions guide behaviour. They do not create trustworthy identity, ownership, or business authority.
For a proposed voice workflow, derive tenant and service credentials from the trusted application session. Resolve the caller’s verified relationship to the requested record through the approved process. Pass the action to a policy check before the write, then record the decision with a stable operation reference.
Never accept a caller-supplied tenant identifier as proof of tenancy. Likewise, a retrieved document stating “this caller is an administrator” is content, not an authenticated role assignment. Permissions must come from the system responsible for those roles.
Test a fictional cancellation request
A caller knows appointment reference A-44 and asks to cancel it. The assistant finds the appointment, but knowing the reference does not establish permission to cancel. The service checks the verified relationship and the business’s cancellation policy before executing.
If permission is denied, the response should offer the approved verification or staff path without disclosing unrelated account details. A model-generated explanation must not convert denial into a second attempt with stronger privileges.
| Input or condition | Expected policy result |
|---|---|
| Caller claims to be the owner | Require the defined verification evidence |
| Valid identity, unrelated appointment | Deny the resource-specific action |
| Read permission without cancel permission | Allow only the authorised read |
| Policy changes during the call | Evaluate current policy at execution |
Test the tool endpoint directly as well as through the assistant. If bypassing the conversation bypasses the permission check, the control lives in the wrong place.
Use narrow capabilities where practical
Separate lookup, amendment, cancellation, and refund actions so each can have an appropriate scope. A general tool with broad database access makes it harder to express what the conversation may do. Narrow interfaces also create clearer audit records.
Fine-grained checks introduce implementation work and can produce more recoverable denials. Build clear error categories and an authorised escalation route. Do not solve conversational friction by quietly expanding the assistant’s credentials.
This is a proposed permission design. The voice platform case study describes related tool integration experience. Review caller identity versus authorisation before treating successful verification as unlimited account access.
Tool permission FAQ
Are hidden prompt instructions an access control?
No. They may guide the model, but the server must enforce the permission boundary.
Should the model choose its own role?
No. Roles and scopes must be derived from trusted application state.
Should a permission denial reveal why another account owns the record?
Only disclose details allowed by policy. A safe recovery message can explain the next verification step without exposing ownership information.