Security and AI Boundaries

Authentication and workspace isolation

Clerk manages authentication. Persisted records are scoped to a ReplyPilot workspace, and owner checks protect workspace deletion. Server credentials are stored as Worker secrets rather than exposed to browser code.

Storage and deletion

Normalized messages, reports, assets, drafts, and source configuration are stored in Neon. Original CSV files use a dedicated R2 bucket rather than the OpenNext cache bucket. Workspace deletion removes the R2 objects before deleting workspace database records through cascading relationships.

• Dedicated upload bucket
• Workspace-prefixed object keys
• Database cascade deletion
• Authentication identity retained separately

AI safety boundary

Deterministic rules identify known privacy, security, fraud, legal, safety, and account-change risks before generation. Red-risk messages remain human-only. Generated FAQs, macros, and drafts require review before export or use.

Connected-source boundary

Gmail access is read-only. Zendesk synchronization excludes agent replies and private notes from customer-demand counts. Approved Zendesk drafts are written as private internal notes rather than automatic public replies. Integration credentials should be encrypted at rest before real customer use.

Current limitations

ReplyPilot does not currently claim SOC 2, ISO 27001, HIPAA, PCI DSS, or another independent certification. Customers should avoid uploading payment card numbers, authentication secrets, medical records, or unrelated sensitive data.