Blog/Automation in practice

The invoice automation process, step by step, and where accuracy is decided

The invoice automation process runs from receipt through capture, validation, coding and approval to posting. What each step does, where accuracy is won or lost, and what to ask of automated invoice processing software before you buy it.

The invoice automation process takes a supplier invoice from the moment it arrives to the moment it is posted in the ledger and ready to pay, in six steps: receipt, capture, validation, coding, approval and posting. Each step can be automated, and each step has its own way of going wrong. Ardent Partners' State of ePayables 2025 put the average cost of processing one invoice at $9.84, the average time at 8.2 days, and the share of invoices that become exceptions at 18.4 percent. The full report found that 35.4 percent of invoices are processed straight through, with no person touching them, which means about two in three still need someone at some step. Where the invoice stops, and why, is decided by a small number of choices in how the process is built.

The invoice automation process, step by step

The process has six steps, and each one passes a record to the next.

  1. Receipt: the invoice arrives as an email attachment, through a supplier portal, as a structured e-invoice or EDI message, or on paper.
  2. Capture: the system reads the header fields (supplier, invoice number, date, amounts, tax, payment terms, PO number) and the line items.
  3. Validation: the captured data is checked against the vendor master, the purchase order and the goods receipt, and against earlier invoices for duplicates.
  4. Coding: the invoice gets its general ledger account, cost center, tax code and any project or entity code.
  5. Approval: the invoice goes to the person with authority for that supplier, amount and cost center.
  6. Posting: the approved invoice is written to the ERP as a payable, scheduled for payment on its due date.

Most of the cost sits in two places. Capture decides whether the data entering the process is correct, and validation decides whether a correct invoice is also one the company should pay. Coding and approval add days when they wait on people, and posting is where a poorly integrated tool creates rekeying at the end of an otherwise automated process.

Receipt looks like the simple step, and it shapes everything after it. Ardent found that only 51.4 percent of invoices are submitted electronically, with the other 48.6 percent arriving by manual means. Even the electronic half is mixed. A structured e-invoice or EDI message arrives as data, with fields already labeled, and needs no capture at all. A PDF attached to an email is electronic only in transit: it is still a picture of a document that has to be read. Paper has to be scanned before it can be read.

The practical design is a single intake that accepts every channel and records where each invoice came from. Suppliers send invoices to the wrong address, send the same invoice twice through two channels, attach three invoices to one email, or put a statement or a credit note in the AP inbox alongside the invoices. Separating documents, classifying each one as an invoice, credit note, statement or something else, and keeping the email that carried it are all part of receipt. A process that gets this step wrong passes unreadable or misclassified documents into capture, and they surface later as exceptions with no obvious cause.

Capture and where accuracy is decided

Capture is where accuracy is decided, and accuracy is measured field by field. A vendor quoting 98 percent accuracy usually means 98 percent of fields read correctly, which sounds high until it compounds. An invoice with fifteen header fields read at 98 percent each, assuming the errors are independent, comes through with every field correct only about 74 percent of the time, so roughly one invoice in four has at least one wrong field. Line items multiply the count further. A useful accuracy figure names the fields, weights them by consequence, and is measured on the company's own invoices.

Not all fields carry the same risk. A wrong invoice date shifts a due date by a few days. A wrong amount, supplier or invoice number can pay the wrong party, pay twice, or block a valid payment, and a wrong bank account sends money somewhere it can't easily be recovered. Good capture reports a confidence for each field and routes low-confidence values to a person, so the reviewer checks the two uncertain fields on an invoice rather than retyping all fifteen. The other half of capture accuracy is the variety of layouts suppliers use, and the share of documents that template-based tools never read reliably; we treat that problem in the long tail of document variability.

Invoices are also untrusted input. A document from outside the company can carry text written to mislead whatever reads it, and software that uses a language model to read invoices has to treat the content as data, never as instructions. We describe the containment for that in prompt injection through untrusted documents.

Validation, duplicates and fraud checks

Validation turns a correctly read invoice into one that is safe to pay. The supplier is matched to a record in the vendor master, and the remit-to bank details on the invoice are compared with the verified details on that record. The invoice is matched to its purchase order on price and to the goods receipt on quantity, within the tolerances finance sets. We walked through that match and its exceptions in three-way match, rebuilt. Invoices with no purchase order go to the requester or budget owner to confirm what was bought.

Duplicate checks need more than an exact match on invoice number. The same invoice comes back as a reminder with a new date, as a scan of the paper copy after the PDF was already paid, or with the invoice number typed as "INV-0042" in one place and "42" in another. Good duplicate detection normalizes invoice numbers and compares supplier, amount and date within a window, and it holds possible duplicates for a person rather than rejecting them.

Fraud checks sit in the same step, because business email compromise, one of the most common forms of payments fraud, works through real supplier relationships. The FBI's Internet Crime Complaint Center recorded 21,442 complaints of business email compromise in its 2024 annual report, with reported losses of $2.77 billion. The IC3 describes the scam as one that targets businesses working with suppliers, using compromised email accounts and social engineering to redirect transfers of funds. The Association for Financial Professionals found in its 2026 Payments Fraud and Control Survey that 76 percent of organizations experienced attempted or actual payments fraud in 2025, and that 58 percent reported check fraud. The control is simple to state: an invoice whose bank details differ from the vendor master, or that arrives with a request to change them, is held until someone confirms the change with the supplier through a contact already on file.

Coding, approval and posting

Coding is where invoices without a purchase order cost the most. A PO invoice inherits its account and cost center from the order. A non-PO invoice has to be coded from scratch, usually by someone who knows how that supplier's invoices were coded last time. Ardent lists invoice coding among the uses of machine learning in AP, and the method is the obvious one: propose the coding the company used for the same supplier and description before, show the history behind the proposal, and let a person confirm it. Coding errors don't stop a payment, which is why they matter. They surface at month end as misstatements in cost centers and accruals.

Approval is the step AP leaders complain about most often. In Ardent's 2025 survey, 49 percent of AP executives listed slow approvals among their top challenges, the most cited, just ahead of a high exception rate at 48 percent. Automation helps most by routing each invoice to the right approver the first time, from the delegation of authority, and by putting the order, the receipt and any variance in front of the approver so the decision takes a minute. Posting then writes the approved invoice to the ERP. The test of posting is whether anything is retyped, since a tool that exports a file someone has to import by hand has moved the manual work to the end of the process.

What to ask of automated invoice processing software

A short list of questions sorts automated invoice processing software better than a feature grid. Ask for field-level accuracy measured on a few hundred of your own invoices, including the worst ones, and ask which fields the figure covers. Ask what the software does with a field it isn't sure of. Ask how it validates against your vendor master and your purchase orders, and what an exception looks like when it reaches a person: the invoice alone, or the invoice with the order, the receipt and the reason it stopped. Ask how duplicates are detected beyond the exact invoice number, and what happens when bank details on an invoice differ from the vendor master. Ask whether it posts to your ERP directly and what it logs, by user, for each change. And ask for the straight-through rate on invoices like yours, against Ardent's average of 35.4 percent and its best-in-class figure of 51.0 percent.

A reasonable counter, answered

A reasonable counter is that e-invoicing will make most of this unnecessary. A structured e-invoice arrives as data, so capture disappears. The trend is real, and Ardent found that 83 percent of AP leaders expected the number of e-invoices they receive to increase in 2025. But the same research found that after two decades of supplier enablement, about half of invoices still arrive in a form someone has to read, and a mid-market company with hundreds of small suppliers will receive PDFs and paper for years. And e-invoicing removes only the capture step. A structured invoice can still carry the wrong price, arrive twice, come from a compromised supplier account, or lack a purchase order, so validation, duplicate checks, coding and approval stay in the process whatever format the invoice arrives in.

Common questions

What are the steps in the invoice automation process?
The invoice automation process has six steps: receipt, capture, validation, coding, approval and posting. Receipt brings the invoice in by email, supplier portal, e-invoice or paper. Capture reads the fields and line items, and validation checks them against the vendor master, the purchase order and the receipt. Coding assigns the general ledger account and cost center, approval routes the invoice to the right person, and posting writes it to the ERP for payment.
How accurate is automated invoice processing?
Accuracy depends on which fields are measured and on how varied the invoices are, so a single headline percentage tells you little. Field-level accuracy compounds across a document: a system that reads each field correctly 98 times in 100 will get at least one field wrong on roughly a quarter of invoices with fifteen fields, assuming the errors are independent. What matters is accuracy on the fields that move money, such as the amount, the supplier, the invoice number and the bank details, and whether the software flags the fields it is unsure of. Ask for accuracy measured on a sample of your own invoices.
What should automated invoice processing software check before an invoice is paid?
It should confirm the supplier exists in the vendor master and that the bank details on the invoice match the verified record. It should match the invoice to its purchase order and receipt within the tolerances the finance team sets. It should look for duplicates, including the same invoice resent with a slightly different number or format. And it should hold any invoice that carries new or changed payment instructions until a person has confirmed them with the supplier through a known contact.

Book a demo

Start with one department

We map how one team’s work runs, show you where the hours go, and automate from there.