Data handling
This page is the inventory security and privacy reviewers ask for: what enters Qefro, what stays in your systems, and what Qefro does not claim.
Legal text: Privacy, Terms. Confirm DPA and subprocessors with Sales.
Short definition (citation-ready)
Qefro stores tenant configuration, workspace knowledge, conversation transcripts, and encrypted integration secrets to operate assistants. Your ERP/CRM remains the system of record for business ledgers. Qefro applies limited pattern-based PII masking on chat messages; it is not a DLP product. Product intent is not to train public foundation models on customer content — confirm that in your agreement.
What Qefro stores (typical SaaS tenant)
| Category | Examples | Isolation |
|---|---|---|
| Account | Users, roles, teams, billing entitlements | Tenant |
| Knowledge | Uploaded docs, website crawls, chunk/vector indexes | Tenant + workspace |
| Conversations | Widget / WhatsApp / portal messages, citations | Tenant + workspace |
| Tools & apps | Tool definitions, encrypted secrets, SDK connections, Marketplace installs | Tenant (+ install) |
| People (Customer Hub) | Channel identities and hub person records when enabled | Tenant |
| Logs | Org audit, identity audit, tool executions | Tenant |
What should stay in your systems
- Orders, invoices, inventory, medical records, payroll ledgers
- Customer passwords, OTP codes, payment card PANs (do not paste them into chat)
- Privileged admin APIs — expose only least-privilege Business Tools
Tool responses may be copied into the conversation the model sees. Design tools so they return the minimum fields required for the answer.
PII masking (honest scope)
Inbound chat messages are passed through scrub_pii before further handling. Today that masks:
- US Social Security Number-like patterns (
123-45-6789) - 13–16 digit sequences that resemble payment cards
Identity audit additionally drops keys named otp, code, token, jwt, password, secret, and authorization from stored detail JSON.
LLM provider requests may send a hashed anonymous end-user id (visitor / WhatsApp contact) — not raw PII — when the provider supports a user field (for example DeepSeek rate-limit isolation). Confirm current providers with Sales.
Models and training
Product intent: customer content powers that tenant’s assistants (retrieval, prompts, tool results). It is not used to train public foundation models on your corpus.
Confirm the contractual wording in your DPA / order form. Model vendors you enable (directly or via Qefro) have their own data-use terms.
Retention and deletion
| Action | Effect |
|---|---|
| Delete a document | Removes the document row and associated vector chunks for that tenant |
| Delete a conversation | Tenant-scoped conversation delete |
| Delete a workspace | Workspace and its scoped content (hard-deleted workspaces no longer count toward entitlements) |
| Delete organization | Owner-driven tenant delete — treat as irreversible; confirm current cascade with support before relying on it for a legal request |
| Privacy / account requests | [email protected] as described on Privacy |
Do not assume infinite log or transcript retention. Export what your policy requires.
Workflow
Classify data before go-live
- List sources — FAQs, policies, ERP fields, tool payloads, WhatsApp bodies.
- Mark public vs internal vs regulated — Only public-safe material in Customer AI workspaces.
- Minimize tool output — Return ids and status, not full customer records, unless required.
- Set retention owners — Who exports conversations and tool logs, and on what cadence.
- Document subprocessors — Request the current list from Sales for your DPA.