The app can read completed payments from a bank statement in two ways. Both end up doing the same thing — recognizing the individual transactions and matching them to the payments you already track — but they differ in where your data travels along the way.
The no-AI path (recommended for more sensitive data)
You upload an export in the ABO/GPC format (commonly available in internet banking, see the import help). The app parses the file entirely by itself, on its own server — nothing from the statement goes anywhere else. The downside: you have to export this format from your bank first, and not everyone has it at hand.
The AI-assisted path (for PDFs)
Almost everyone has a PDF statement right in their inbox, but reading it automatically is harder — the layout differs from bank to bank. So the app:
- Extracts the text from the PDF right in your browser (no file is sent anywhere).
- Sends that text (not the PDF) to Anthropic's Claude API, which recognizes it as a list of transactions and suggests a category/possible match.
- The decision whether a payment is matched is made by the app itself — by comparing the account number and variable symbol (or the amount and date). The AI suggestion is only an aid; it is never used directly as "this is paid."
What we know about data processing at Anthropic
- Data sent through the API is not used to train models by default.
- With the standard API, requests/responses are kept only briefly (days), then deleted automatically.
- Processing runs on infrastructure in the US, not in the EU.
If that doesn't suit you, use the no-AI path above — both do the same thing, just by a different route.
What the app does with the data after processing
The app does not permanently store the uploaded file or the raw statement text — only the resulting structured details of the individual transactions (date, amount, counterparty account, symbols), needed so that next time it can recognize the same payment.