3.8 FTE of keying returned - with 76% of documents now posting themselves

RESULTS

Measured over the first 90 days after go-live.

3.8 FTE returned
604 hours of keying a month came off the team, redeployed onto supplier queries and month-end rather than cut.
76.5% straight through
Read, validated and posted with nobody opening them. Invoices reach 87.2%; KYC packets 43.3%, and never higher by design.
Cost per document: $6.23 -> $0.51
A 92% reduction on the highest-volume task in the centre, model spend included.
Arrival to posted in 54 seconds
For a matched invoice, against a previous median of the next working day - keying ran as a morning batch.
Nil variance across five systems
Every posted record is checked back against what the destination actually holds, daily, since go-live.

The problem

Extraction was never the hard part. Keying was.

Five companies, five formats, one keyboard
4,600 documents a month arrived by email, SFTP, portal, scanner and API, and every one was retyped into NetSuite, Guidewire, Fenergo or CargoWise by hand.
11 minutes each, blended
843 hours a month - 5.3 full-time equivalents - spent moving data that already existed from one screen to another.
Tools that stop at extraction do not help
They had trialled OCR before. It read the document accurately and left the keying exactly where it was, because nothing wrote the record.

What we shipped

A pipeline that finishes the job rather than stopping at the read.

One intake for every channel
Email, SFTP, portal, scan and API, whatever the document type.
Extraction with a confidence score per field
Not per document - the distinction is what makes routing possible.
Three-way match on invoices
Against the purchase order and the goods receipt, within a 2% or $50 tolerance.
Blocking rules per type
Duplicate detection, sanctions screening, identity expiry, policy in force.
The write itself
Into five different systems of record, with the payload and the returned record id kept against every document.

607 documents a month still stop for a person. Most are a PO mismatch or a missing goods receipt, which is a data problem upstream rather than an extraction failure, and every sanctions hit goes to a human whatever the confidence score. Six months from first call to go-live; the engagement is ongoing.

Five operating companies sent invoices, claims, KYC packets, purchase orders and shipping documents into one shared-services team, and a person keyed every one of them into a different system. We built a pipeline that reads whatever arrives, validates it against the rules that matter, and writes the record itself. 76.5% now go from arrival to posted with nobody opening them, and 604 hours of keying a month went back to the business.