ML-driven cash application

The client builds an accounts-receivable automation product that finance teams use to collect and reconcile customer payments.

When a customer pays, the money arrives with a remittance email that may or may not say which invoices it covers. The detail sits in free text, an attachment, a spreadsheet or a scan, and finance teams were matching each payment to its invoices by hand.

We built the pipeline that reads those emails. It takes in the message and its attachments, parses the remittance advice in whatever format it comes, and hands the result to models that extract the entities: amounts, references, dates, payer. Further down the pipeline, business logic uses the semantic and search layers to allocate the funds against the client's open documents: invoices, credits and adjustments.

Today 60% of payments are allocated automatically: $2.5bn a year in customer payments matched to invoices without a person touching them.