AI bank-payment reconciliation for a hosting provider

The problem

A hosting provider issues smart invoices to its clients in Bitrix24. When a payment arrived, a manager had to open the bank's incoming-payment email by hand, find the matching invoice, cross-check the counterparty, amount and payment purpose — and manually move the invoice to the «Paid» stage. Dozens of payments, each one a manual reconciliation. A payment could sit unprocessed for hours: the client had paid, but the service was formally still not activated.

Before

• The manager read the bank's incoming-payment emails by hand and looked up the matching smart invoice.
• The «counterparty + amount + purpose» check was done by eye, with the risk of attaching a payment to the wrong invoice.
• Between the actual payment and the «Paid» status in the CRM — a lag of several hours.

What we built

• A daemon that holds a 24/7 connection to the mailbox and instantly catches the bank's incoming-payment emails — no hourly polling.
• The counterparty, tax ID, amount and payment purpose are extracted from the email. The email text and the list of open smart invoices go to AI (Claude), which matches the payment to an invoice and returns a result with a confidence level and a rationale.
• On a confident match the invoice is moved to «Paid» automatically — no human involved. When in doubt the payment is not attached blindly; it goes to the manager in Telegram for a manual check.
• Fault tolerance: the daemon restarts itself on failure, reconnects on a dropped connection and separately monitors its own availability, alerting via Telegram.
• An engineering catch: AI providers block requests from Belarusian IPs, so calls to the model go through a self-hosted relay proxy located outside the country.

Results

• Manual payment reconciliation as a process is gone: on a confident match the invoice closes itself within seconds of the bank's email.
• Ambiguous payments are neither lost nor attached at random — a human decides on them deliberately.
• The payment status in the CRM no longer lags reality: the client pays and the service is provisioned almost immediately.
• The manager handles only the exceptions, not the routine confirmation of every payment.