The AP automation stack: capture, matching, workflow, payment
The AP automation stack is the sequence of four layers that carries an invoice from arrival to settlement: capture, which converts documents into structured data; matching, which verifies that data against purchase orders and receipts; workflow, which routes whatever fails verification to someone who can resolve it; and payment, which executes and records the settlement. The layers run in series, so each one inherits the quality of the layer before it. That single fact explains most of what goes right and wrong in AP automation.
Capture feeds every layer above it
Capture ingests invoices from email, supplier portals, EDI, and paper, then extracts header and line-level fields into a structured record. Modern extraction handles clean, consistent documents well and struggles with the messy ones – handwritten freight invoices, multi-page utility bills, line items that describe the same goods differently than the PO does. The struggle matters because a field extracted wrong does its damage two layers later, when a match fails for reasons nobody can see in the invoice image. Our article on what OCR and invoice capture actually solve covers where extraction ends and interpretation begins.
Matching decides how much work is left
Matching compares the captured invoice to the purchase order and, in a three-way match, to the receiving record. When all three agree within tolerance, the invoice posts without anyone touching it. Every disagreement – a price variance, a quantity gap, a unit-of-measure mismatch, a missing receipt – becomes an exception. Matching is the layer that determines how much of AP is automated in practice, because everything it fails turns into manual work for the layer above.
Workflow inherits every exception matching produces
Workflow is the routing and approval layer: it assigns exceptions to owners, chases approvals, and enforces delegation-of-authority rules. It also carries the coding burden for invoices that arrive without a PO – see our guide to GL coding. This is where most stacks quietly turn back into manual work. The workflow tool routes an exception to a queue, and a person then investigates it by hand: opening the ERP, checking the receiving system, emailing the supplier, reading the contract. The routing is automated while the resolution stays manual, and that gap is the main subject of our article on where AP automation projects stall.
Payment is the easiest layer because the hard decisions happen upstream
Payment execution – scheduling runs, selecting methods, capturing early-payment discounts, producing remittance files – is the most mature layer, and banks, ERPs, and payment providers all handle it competently. An approved, coded, matched invoice is easy to pay. The reason payment feels solved is that every judgment call was made before the invoice got there. Teams that shop for payment tools to fix AP throughput usually discover the constraint sits two layers down.
The weakest layer sets the throughput of the whole stack
Because the layers run in series, the stack clears invoices at the pace of its most manual layer, and in most AP departments that is exception resolution: the investigation work sitting between matching and payment. Capture and payment are mature, and matching logic is deterministic. Exceptions are where automation has historically stopped, because resolving one requires reasoning across the ERP, the receiving system, supplier emails, and spreadsheets at once – work that rule-based RPA cannot do and that an ERP suite cannot see beyond its own tables. A useful evaluation of any AP automation stack starts with a single question: what happens to the invoices that fail the match?
Every other layer of this stack has a mature vendor market; the work sitting between matching and payment is the part the industry has left to your analysts. Fragment builds agents for exactly that layer – they resolve invoice exceptions, duplicates, and coding gaps by reasoning across your existing systems in place – the ERP, the receiving system, supplier emails, spreadsheets – running as autonomously as you trust them to, with human breakpoints wherever you want them. Browse the workflows Fragment covers or book a demo to see one run on your own exception queue.