Security at AI Matrix

Purchase orders, invoices and payment data are sensitive. Here is exactly how we protect them — in plain English, without overstating anything.

Every customer's data is separate

  • Every record carries its workspace ID, and the database itself refuses to return or change rows for any other workspace (row-level security) — not just the app.
  • Files are stored privately under the workspace's own folder and are only shown through the app or short-lived links.
  • Automated tests check that reading or writing across workspaces is refused.

Bank details are never stored

  • Bank account numbers read from invoices are turned into a keyed one-way fingerprint. Only the fingerprint and the last four digits are kept.
  • A different account on an invoice is always flagged, and invoices with a bank change can never be approved automatically.
  • Changing a vendor's bank details needs an admin, the verification method and the name of the person who confirmed it — and is recorded in the audit trail.

People make the decisions

  • AI Matrix raises risk signals; it doesn't pay anyone and doesn't change your accounting records.
  • Approval limits by amount decide who can approve. Blocking exceptions need an explicit override.
  • Automatic approval is off unless you switch it on, and only applies to clean invoices under your limit from vendors with verified bank details.

A complete, append-only audit trail

  • Every document received, every check, every time someone opens an invoice and every decision is recorded with who and when.
  • The trail is append-only in the database: nobody — including workspace owners and our own application — can edit or delete entries.
  • Export it to CSV at any time for your auditors.

How AI is used

  • Documents are read with Anthropic's Claude API. Anthropic states that it doesn't use inputs or outputs from its commercial API to train models by default.
  • Text inside documents and emails is treated as data, never as instructions. The model can only return a structured record, which is validated before it's used.
  • Every field comes with a confidence score; low-confidence reads are flagged for a person.

Access control

  • Five roles — owner, admin, approver, reviewer and viewer — each limited to what they need.
  • Sign-in by email link or password, with verified email addresses. Invitations are limited and logged.
  • Forms are protected against cross-site requests, and outgoing webhooks are signed and can only reach public HTTPS addresses.

Email intake

  • Each workspace has its own forwarding address. Only PDF, image, CSV and Excel attachments up to 25 MB are read.
  • Links in emails are never opened, and each message is processed once even if it's delivered twice.

Encryption and hosting

  • All traffic is encrypted with TLS. Data is stored with Supabase (PostgreSQL and file storage), which encrypts data at rest.
  • The application runs on Vercel. Payments are handled by Stripe — card details never reach our servers.

Sub-processors

SupabaseDatabase, authentication, file storage
VercelApplication hosting
AnthropicDocument reading (Claude API)
StripeSubscription billing
PostmarkInbound email

Reviews and questionnaires

We're happy to complete your security questionnaire, walk your IT team through the architecture and sign a data processing agreement. We don't claim certifications we don't hold.

Found a vulnerability? Email hello@matrixaicenter.com and we'll respond quickly.

Request a security review

AI Matrix is a product of Matrix AI Center LLC.