Cutting loan file processing from six days to under two

Kaligandaki Finance processes around 400 loan applications a month. Each one arrived as a bundle of scanned documents (identity papers, proof of income and collateral documents) and was typed into the core banking system by hand. The average turnaround was six working days, and almost all of that delay came from data entry.

Two earlier attempts at automation had failed. Both tried to automate the lending decision as well as the data entry, which the credit committee refused to accept, and with good reason.

We deliberately limited the automation to extracting data. The system identifies each type of page, extracts the fields with a confidence score, and sends anything uncertain to a member of staff, who sees the original scan alongside the extracted values.

How it works

• A service that accepts scans from branch scanners and from mobile phones.
• Page classification and field extraction, using models trained on the layout of the standard forms.
• A review queue where staff correct uncertain fields. Their corrections are used to measure and improve accuracy.
• Records written directly into the existing core banking system through its API, with a full audit log.

Lending decisions were left entirely to people. That was deliberate, and it is what made the project acceptable to the committee that had blocked the earlier attempts.

The average processing time fell from six working days to 1.6. Staff who had been typing in data moved to the review queue, where each person now handles about four times as many files. At handover, the system extracted 96.4% of fields correctly before review, and effectively 100% after it.

The client has since extended the same system to account-opening documents.