Touchless Invoice Processing: What 94% Straight-Through Actually Takes
Modern document AI reads a vendor invoice better than a tired human at 4:45 on a Friday — that part is genuinely solved. So why do so many “automated AP” projects still route half their volume to manual review? Because extraction is maybe a third of the system. The touchless rate lives in what happens after the reading.
Anatomy of a 94% system
- Extraction: the model pulls header, line items, tax, and terms from any layout — PDF, scan, or photographed paper.
- Validation: totals recompute, vendors match the master file, duplicates get caught before they post.
- Matching: two- or three-way match against POs and receipts, with tolerance rules mirroring your actual policy.
- Coding & routing: GL codes assigned from learned history, approvals routed by amount and department automatically.
- Exceptions: the honest 6% land in a queue with the reason attached — humans handle judgment, never data entry.
Where projects actually stall
The failure is rarely the AI. It's vendor masters full of duplicates, PO policies that exist on paper but not in practice, and tolerance rules nobody has written down. Automating a process forces you to finally define it — that definition work is unglamorous, and it's also most of the value. The companies that hit 94% treated the cleanup as part of the project, not a surprise.
The payoff profile
A mid-size company running 600 invoices a week recovers roughly 12,000 hours a year at 94% touchless — plus early-payment discounts captured, late fees eliminated, and an AP team doing analysis instead of typing. The ROI usually lands within two quarters, and unlike headcount, it doesn't call in sick during month-end close.