Permission to read a source does not automatically justify every answer generated from it. The task, recipient, and disclosure policy can impose an additional boundary on how authorised information is used.
Access and permitted output are distinct decisions
Source access determines whether a trusted principal may retrieve a resource. Answer permission determines whether the requested use and recipient are allowed to receive the resulting content under the business policy.
For a proposed workflow, keep user identity, source audience, task purpose, intended recipient, and output scope explicit. Enforce retrieval permissions first, then apply the approved disclosure rules to the answer and any export or send operation.
Do not rely on this second layer to excuse overbroad retrieval. If the source is not needed for the task, avoid supplying it. Data minimisation and permission enforcement remain complementary controls.
Test a fictional internal-to-customer reply
A support employee can read internal incident notes to investigate a problem. They ask the assistant to draft a customer reply. The reply may explain approved facts and next steps, but should not expose private detection logic or unrelated customer details from the notes.
Another request asks for an internal investigation summary for an authorised reviewer. That can have a different output scope. The source access is similar, while the permitted recipient and purpose differ.
| Task and audience | Output boundary |
|---|---|
| Internal authorised review | Relevant permitted evidence |
| Customer-facing reply | Approved external disclosure only |
| Broad company announcement | Publication authority and minimisation |
| Bulk export | Separate scope and export permission |
| Unclear recipient | Clarify before protected delivery |
Use synthetic restricted markers and evaluate the exact delivered artifact, including attachments and citations. A safe draft can become unsafe through later template expansion or an unreviewed send step.
Preserve truthful explanations
Output restrictions should not produce invented stories. If details cannot be disclosed, use the approved accurate explanation and available next step. Keep speculation, recommendation, approval, and completed action clearly distinguished.
Two-stage policy adds complexity and review work. Make the rules concrete enough to test rather than asking the model to decide vague confidentiality on its own. The trusted application should enforce consequential delivery scope and retain an accountable owner.
This answer-boundary design is proposed. The voice integration case study describes related support workflows. Pair it with internal notes and external replies.
Source versus answer FAQ
Can an employee publish everything they can read?
No. Reading, external disclosure, publication, and export can have different authority requirements.
Does output review replace retrieval permissions?
No. Enforce the source boundary before generation too.
What should a test compare?
Hold the source scope constant while varying task and recipient, then verify permitted and forbidden claims and delivery effects.