Data Flow & Security Overview
A plain-language summary for prospective customers and their security/procurement teams. Not a substitute for the Privacy Policy or Data Processing Addendum, which govern in case of any conflict.
1. What data LangevinAI touches
When you use LangevinAI, you upload a general ledger export (Excel/CSV) and, optionally, supporting documents (PDF, Word, .eml). These files typically contain:
- Account names and balances
- Transaction-level detail (dates, descriptions, amounts, vendor/customer names)
- No payment card or bank account credentials are collected by LangevinAI itself
2. How your data moves through the system
Shared-proxy mode (Solo, Pro, and Business plans)
- You upload a file in your browser. Parsing (header detection, format normalization) happens client-side in JavaScript — the raw file is not sent anywhere just to be parsed into a table.
- When you request AI classification, commentary, or summaries, the relevant account names, sampled transaction rows, and document excerpts are sent from your browser to our Vercel serverless function.
- The function verifies your identity (Clerk-issued session token) and your plan entitlement (Supabase), applies model and fair-use limits, then forwards the request to the Anthropic API over TLS, using our Anthropic API key.
- Anthropic returns the classification/commentary result, which is passed back to your browser and rendered. It is not written to a database on our side.
- Your file itself is not stored on our servers — the working data lives in your browser session for the duration of your use of the app.
BYOK option (Bring Your Own Key, available on any plan)
- Same client-side parsing as above.
- Your personal Anthropic API key is stored only in your browser's local storage — it is never transmitted to, or stored on, our servers.
- AI requests go directly from your browser to the Anthropic API using your key. We never see the request or response content in this mode.
3. What we store, and where
| Data | Stored? | Where |
|---|
| Uploaded GL/budget/support files | No | Not persisted server-side; held in-browser only |
| Account classifications, AI commentary | No | Held in browser session only, regenerated as needed |
| Your API key (BYOK) | No (by us) | Browser localStorage only |
| Account/plan/billing info | Yes | Supabase (plan status, usage counters), Clerk (authentication), Stripe (payment/subscription records) |
| Login credentials | Not by us directly | Managed entirely by Clerk (we never see passwords) |
4. Subprocessors
| Subprocessor | Role | Data it may see |
|---|
| Anthropic | AI model inference (classification, commentary, summaries) | GL account names, sampled transactions, supporting-document excerpts, per-request only — not retained by LangevinAI |
| Vercel | Hosting, serverless function execution | Request metadata; transient in-memory access to data in transit through the proxy |
| Clerk | Authentication (sign-in/sign-up, session tokens) | Name, email, authentication metadata |
| Supabase | Plan/entitlement storage | Clerk user ID, plan tier, entity count, usage counters |
| Stripe | Billing and subscription management | Payment details, billing contact info |
We do not use customer financial data to train AI models, and we do not sell customer data to third parties. See the Privacy Policy for our data-retention commitments in full.
5. Why this matters for your review
- No server-side data lake of customer financials — there's nothing to breach because there's very little sitting still.
- BYOK mode gives security-sensitive customers a path where their financial data never touches our infrastructure at all.
- Single-file architecture means the entire client application is auditable in one file — no obscured data paths across dozens of services.
If your security team needs a completed vendor security questionnaire or a signed Data Processing Addendum, contact us at contact@langevinai.io.
