Your client data, protected like ours.

• All customer data is stored in Postgres on Supabase (AWS us-east-1), encrypted at the volume level with AES-256. The one thing kept outside it is abuse-prevention counters, held in Redis in the same region — a request tally keyed by your internet address or a scrambled email, kept at most an hour, and never bookkeeping data.
• Every query is scoped to your firm in our server code, and every table is locked with row-level security so the public database API can read nothing at all. The app itself connects as a trusted server role, so firm scoping is enforced in application code, and our integration tests seed a second firm to check data doesn't cross between firms.
• Bank and QuickBooks Online access tokens, other connection credentials and two-step secrets are additionally encrypted by our code with AES-256-GCM before being written to the database. Other fields — names, bank descriptions, emails — rely on the volume encryption above.
• Backups are encrypted by Supabase and taken daily; each one rolls off after 7 days. Firm deletion purges every row scoped to that firm, so it is gone from backups a week later.

TLS 1.2+ enforced end-to-end via Vercel's edge. HSTS is set on every response with a 2-year max-age. Outbound transactional email is signed with SPF and DKIM through Resend.

The categorization model (OpenAI GPT-4o-mini, with Claude Sonnet as a configurable alternative) receives only the bank line's vendor name and description, its amount and whether the money came in or went out, its memo when it has one, and the names of that location's categories that can take that money, so it can only pick one of yours. While “Use past decisions” is on, it also receives up to eight lines your firm already categorized that look most like it — vendor, description and the category chosen — from any of your locations, when that category is on the list. To find those look-alikes, OpenAI's embedding model reads each line's vendor, description, amount and the location's business type. The date is never sent, and neither are location, firm or client names. It never sees balances, payroll details, client contact info, or any data outside the categorization context. Under OpenAI's API terms this data is not used to train its models; OpenAI may keep it for up to 30 days to watch for abuse, then deletes it. We keep no log of prompts — only the suggested category, its confidence and its reason, on the line itself.

We never see them. Bank connections go through Plaid via OAuth. The token LedgerInbox holds can pull transactions but cannot move money, change passwords, or read non-bank data. You can revoke it at any time from Settings → Bank connections or directly at your bank's portal.

• Sell, share, or analyze customer data for any purpose other than running the service.
• Train or fine-tune AI models on customer data.
• Send marketing emails to your clients.
• Look at a firm's books except to fix a problem you report, keep the service running and secure, or comply with the law. Only the people who run LedgerInbox — today, its owner — can reach production systems.

Companies that process customer data for us — all in the United States. How long each keeps it is on the privacy page.

Where your data comes from, when you connect it (each under its own privacy policy):

If you find a security issue, please email support@ledgerinbox.com. We'll acknowledge your report and work with you on a coordinated disclosure timeline. We don't run a paid bug bounty program yet — when we do, this page will say so with the published scope and payout table.