AI support tools should operate within the current customer’s authorised scope. A model-generated account identifier must not decide which tenant or customer records the backend is willing to expose or change.
Tool scope comes from trusted application context
Current-account tool scoping binds an operation to the verified requester relationship and permitted resource set. The model can propose a relevant record reference, but trusted services determine whether that resource belongs to the allowed scope.
For a proposed executor, derive tenant and service identity from authenticated configuration. Resolve customer relationship through the approved verification process. Validate every action against current permissions and resource ownership before reading or mutating protected data.
Do not assume a successful initial verification grants unlimited session access. Roles, delegation, and resources can differ by action, and some consequential operations require additional checks.
Test a fictional cross-customer prompt
A customer asks the support assistant to inspect another account, claiming it belongs to a colleague. The agent may capture the request or explain the verification path, but the tool service must deny unauthorised access even if the model submits the other identifier.
A second fixture gives the caller a legitimate relationship to two accounts. The system should use the explicitly selected permitted context, not silently choose the first lookup result. Scoped access should support valid work without broad credentials.
| Tool request | Expected boundary |
|---|---|
| Permitted current resource | Allowed action under policy |
| Another tenant’s identifier | Denied without private disclosure |
| Caller claims administrator status | No role change from untrusted text |
| Delegation revoked mid-session | Current permission rechecked |
| General public lookup | Separate from protected account tools |
Test the endpoint directly with tampered identifiers and inspect returned metadata as well as writes. A denial that reveals private account names can still violate the intended boundary.
Keep handoff context accurate
When staff take over, preserve the verification level and permitted scope. Do not make them assume the AI completed checks that never occurred. A concise record can show which actions remain unavailable and why.
Strict scoping adds implementation work and may require clarification for legitimate multi-account users. Provide a clear authorised switching or delegation process. Avoid solving friction by letting the model choose arbitrary tenant context.
This tool design is proposed. The StackFlight case study describes private account workflows without claiming this support executor. Pair it with server-side tool permission checks.
Current-account tools FAQ
Can the prompt alone restrict account access?
No. The backend must enforce scope for every request.
Does knowing another account ID authorise access?
No. Resource identifiers are references, not proof of relationship.
What should an access error reveal?
Only safe information needed for the approved recovery path, without disclosing protected account details.