How it works
Bank statements are extracted to normalized JSON at about 95% field accuracy, deterministic knockout rules with confidence thresholds decide which files qualify, and qualified files start a sub 60 second outreach sequence in Salesforce. Edge cases go to a one click manual queue with reason codes, and underwriters keep the final say on overrides and pricing.
What triggers it
A merchant submits bank statements with an application.
Step by step
- 1OCR and an LLM extract each statement into normalized JSON
- 2A rules engine applies knockout rules with confidence thresholds
- 3Qualified files are pushed to Salesforce through webhooks
- 4Salesforce starts outreach within 60 seconds
- 5Edge cases land in a manual queue with reason codes and a full audit trail
Why an off the shelf connector was not enough
Salesforce stores the deal but cannot read a bank statement. The underwriting rules are specific to the lender, so the extraction and decision layer had to be built around them.