Merdial holds recordings of other people’s conversations. That is about as sensitive as business data gets, and the architecture is built around one assumption: that application code will eventually have a bug, so the isolation between tenants must not depend on application code being correct.
1. Tenant isolation
Every table that holds customer data carries both an organisation and a department identifier, and access is enforced by row-level security policies in the database — evaluated by Postgres on every query, against the authenticated user’s own token. A query written incorrectly in the front end returns no rows rather than someone else’s.
- There is exactly one database client in the application. A second one is the usual way a policy gets bypassed, so the codebase forbids it.
- The department, not just the organisation, is the isolation boundary: an agency running five client teams gets five isolated data sets inside one account.
- Elevated service-role credentials exist only server-side, inside edge functions. They are never shipped to a browser or a desktop build.
- Client-portal users see only records their own campaigns produced, and only after a human has approved them.
2. Encryption
- In transit: TLS 1.2+ everywhere, including the websocket carrying live call audio. HSTS on the web surfaces.
- At rest: databases, object storage and backups are encrypted with AES-256 by the managed platform.
- Secrets — vendor API keys, webhook signing secrets — live in a controlled server-side store whose read policy denies them to application clients by construction, not by convention. Rotation is a single operation across every function.
3. Authentication and access
- Named accounts with hashed credentials; no shared logins.
- Brute-force protection: repeated failed sign-ins from an address auto-block it, and administrators can block an address permanently. This is why the sign-in flow looks up the device’s public IP — see the Privacy Policy.
- Permissions are a per-page grant list, not a coarse role. A member’s effective access is the intersection of what the organisation has enabled and what they have been granted individually.
- Ghost mode lets an owner view the product as a member. It is display-only, and an action taken in it is never attributed to the target user.
- Every edge function verifies the caller’s token and re-checks role scope server-side. There are no unauthenticated endpoints that touch tenant data.
- Privileged and administrative actions are written to an audit log with a server-derived source address, not a client-supplied one.
4. What reaches an AI model, and what does not
Redaction happens before the prompt is built, not after
Payment-card patterns are stripped from transcripts before any generative model call. The redaction step is not skippable to simplify a prompt — it sits in the shared path every AI feature goes through.- Prompts carry the minimum context a task needs: the relevant transcript span and the rubric or criteria being applied. Not the contact database, not the account history, not another department’s data.
- Model providers are contracted on no-training terms for the API content we send them. Merdial improves its OWN models on de-identified transcripts — see Privacy §5 — and the vendors are named at Sub-processors.
- Credentials, tokens and payment instruments are never included in a prompt.
Merdial is not a cardholder data environment
The redaction layer is a safety net for speech that happens to contain a card number, not a licence to process payments through the platform. Do not put card data into Merdial — see Acceptable Use §4.5. Backup, recovery and retention
| Control | Detail |
|---|---|
| Backups | Automated daily, encrypted at rest. |
| Point-in-time recovery | Rolling 35-day window. |
| Recovery targets | RPO 5 minutes, RTO 4 hours. |
| Deletion | Customer instruction is honoured on live data immediately; backups age out within the 35-day window. |
| After termination | 30-day export window, then deletion or irreversible anonymisation within 90 days. |
6. How changes reach production
- Version-controlled, reviewed changes. No direct edits to production.
- A typed codebase with an automated test suite, plus a pre-ship security gate for any change to a server-side function — it checks token verification, role scoping, redaction before generative calls, and that no elevated credential can leak to a client.
- Row-level security is re-verified from the outside after any migration: an unauthenticated client is pointed at every tenant-scoped table and must be refused by all of them.
- Dependencies are monitored for known vulnerabilities and patched on a risk basis.
7. Compliance posture — stated honestly
What we hold, and what we do not
Merdial does not currently hold a SOC 2 Type II report or ISO 27001 certification. We are an early-stage company and we would rather tell you that than let you infer otherwise from a page full of badges.
What we do have: a GDPR-compliant Data Processing Agreement with Standard Contractual Clauses, a published sub-processor list with 30-day change notice, the technical controls described above, and a completed security questionnaire we will send on request. Formal audit is on the roadmap; ask team@merdial.com where it stands before you build a procurement timeline around it.
8. Incident response
- A confirmed personal-data breach affecting customer data is notified to affected customers within 48 hours of us becoming aware — nature, scope, likely consequences, and what we are doing. See DPA §8.
- Merdial notifies the customer, not the customer’s data subjects or a supervisory authority, unless the customer asks us to.
- A written post-incident review follows within 5 business days for a critical incident.
9. Reporting a vulnerability
Email team@merdial.com with enough detail to reproduce. We acknowledge within 2 business days, keep you updated, and will credit you publicly if you want that.
- Please do not access, modify or delete data that is not yours, degrade the service, or use social engineering against our staff or vendors.
- Test against your own account. If you need a clean environment to test in, ask and we will give you one.
- We will not pursue legal action against research conducted in good faith within these bounds.
Contact
Security team: team@merdial.com. This overview is current as of 6 September 2026.