A white-label dashboard can show the correct client name while its API returns another client's calls. Interface branding is not tenant isolation.
For an agency operating several AI receptionists, client boundaries must apply to records, files, search, exports, background work, and administrative access.
Bind the client boundary to verified identity
Resolve the authenticated user's permitted client or organisation on the server. A browser-supplied client identifier can select a resource, but it cannot establish permission to read it.
AWS's PostgreSQL tenant-isolation article illustrates one database-level approach. Application checks and database policies need to fit the actual connection and role design; enabling a feature does not automatically protect every access path.
Follow one call record through every surface
For two fictional clients, create distinct calls, transcripts, and recordings. Then test each surface as a user from only one client:
| Surface | Isolation check |
|---|---|
| Dashboard list | Only authorised client records appear |
| Direct call URL | Changing the record ID does not expose another client |
| Recording download | File access is checked independently |
| Export | The dataset stays within the authorised boundary |
| Search and summaries | Results and caches do not cross clients |
Do not stop at the normal navigation path. Test direct API requests and stale links after a user loses access.
Keep background jobs and caches client-aware
A scheduled report or summary job often runs without a browser session. It still needs a trustworthy client context, limited access, and output destination.
Cache keys and invalidation rules need the relevant boundary as well. Two clients asking for “recent calls” must not share a response merely because their query text matches.
For stored recordings, avoid treating an unguessable file path as the entire access policy. Use the storage system's supported controls and authorised delivery mechanism. Consider how links expire and what happens after offboarding.
Make agency support access explicit
Support staff may need cross-client visibility, but that is a separate role to authorise and audit. Do not give every agency user broad access because it simplifies debugging.
Record sensitive administrative actions and remove access when staff or clients leave. A client export should not include internal notes about other clients or shared platform configuration.
The trade-off between separate infrastructure and shared infrastructure depends on risk, cost, and operating capacity. Shared systems can support isolation, but the boundary must be demonstrated rather than inferred from a dashboard screenshot.
The Next.js permission guide covers one application entry point. The agency onboarding article covers assigning client ownership and access.
Client-isolation questions
Is a client filter in the interface sufficient?
No. Enforce permission at trusted server and data boundaries.
Are recordings protected by the same check as call rows?
Not necessarily. Test storage and download access separately.
What should the first release test contain?
Two clients, direct record access, file access, and revoked membership. Use that matrix in an agency platform review.