A shared AI runtime must retrieve the right client's business facts. Similar service names and copied documents make it easy to answer one caller with another company's prices, hours, or contact routes.
Separate reusable conversational policy from client knowledge. Store source ownership, authorized client scope, effective dates, review status, and purpose. Keep professional instructions and consequential commitments within their approved content boundary.
Enforce scope before retrieval reaches the model
Filter or isolate the knowledge store using verified client context and enforce authorized access to source documents. Do not retrieve broadly and ask the model to ignore the wrong client's passages afterward.
Caches, embeddings, attachments, summaries, and background indexing also need scope. A correctly filtered main query can still return stale or cross-client content through a shared cache key or inherited document reference.
Work through a fictional fee collision
Two clients offer a service with the same name but different assessment fees. Client A's call asks about price. The retrieval path must use A's approved current source and preserve its conditions.
If A's source is missing, the assistant should route pricing to staff rather than borrow Client B's answer. The source reference and active knowledge version let operators explain where the stated fee came from.
| Knowledge case | Required control |
|---|---|
| Same topic across clients | Authorized source scope |
| Missing local fact | Honest unknown and owner |
| Expired document | Freshness policy applied |
| Cached answer | Client and version context retained |
| Source update fails | Visible indexing or review gap |
Assign content maintenance ownership
Someone must approve hours, prices, coverage, promotions, and contacts. Technical indexing cannot decide whether those facts remain valid. Record update status and test affected questions after a material change.
Shared infrastructure lowers repeated maintenance but raises the cost of a scope error. Separate stores may simplify access reasoning while adding provisioning work. Choose enforceable boundaries and include negative tests rather than relying on naming conventions alone.
Review client isolation and shared prompt versions for the surrounding runtime and release context.
Questions about shared knowledge
Can every client document live in one index?
Only if the architecture enforces authorized retrieval and lifecycle boundaries appropriately. The storage arrangement alone does not establish isolation.
Should a missing answer use another client's example?
No automatic cross-client fallback should supply business facts. Use a staff-review route within the correct scope.
What should answer traces retain?
Client context, source references, knowledge version, and relevant uncertainty under the data policy, without unnecessary copies of sensitive documents.