Invoice Processing Automation, Without the Vendor Gloss

What you are actually buying

“Invoice automation” covers four separate jobs, and vendors blur them. Capture pulls the data off the document. Matching reconciles it against a purchase order and a delivery. Approval routes it to whoever must say yes. Posting writes it into the accounting system.

Capture is close to solved. Matching and approval are where the money and the difficulty are, because they encode how your business actually makes decisions — and no two businesses do that the same way.

Capture: better than it was, still not magic

Extraction from PDFs and scans is genuinely good now. On typical supplier invoices, field-level accuracy of 90–95% straight out of the box is a fair expectation, higher once the system has learned your regular suppliers' layouts.

The remaining errors cluster predictably: multi-page invoices where line items span the break, credit notes formatted like invoices, handwritten annotations, and suppliers who redesign their template without warning. None are exotic; all need a review path.

Ask any vendor for accuracy on your ten most awkward invoices, not their demo set. The gap between the two numbers is the real project risk.

Matching is where the savings live

Three-way matching — invoice against purchase order against goods received — is the step that stops you paying for things you did not get. Done by hand it is slow, so in practice it gets done thoroughly for large invoices and skipped for small ones, which is precisely backwards from a fraud perspective.

• Exact match: post automatically, no human involved.
• Within tolerance (say 2% or £5 on price, small quantity variance): post with a log entry.
• Outside tolerance: route to a person with the discrepancy highlighted, not the whole document dumped in an inbox.
• No purchase order: route by category and value to whoever owns that spend.

Getting the tolerances right is a business conversation, not a technical one. Set them too tight and you have recreated the manual process; too loose and the control is theatre.

Approval routing: encode the real rules, not the policy document

Every business has a written approval policy and an actual one. The actual one includes “anything from this supplier goes to Dave” and “if it is under £200 nobody really checks”. Automate the real rules or people will route around the system within a month.

Build in escalation from day one: an approval sitting for five days should chase, and after ten it should escalate. Most late-payment problems in mid-sized businesses are not cash flow, they are an invoice sitting in one person's inbox.

What it costs, and what it saves

Manual cost per invoice includes handling, keying, chasing approvals and fixing errors. Below roughly 200 invoices a month the arithmetic usually says wait, unless error costs or audit requirements dominate.

The exceptions that decide the project

Write these down before anyone quotes, because they are the difference between a smooth rollout and a stalled one.

1. Credit notes and negative amounts
2. Multi-currency invoices and the rate you apply
3. Partial deliveries and partial invoicing
4. Suppliers who invoice in a completely different unit from the one you order in
5. Duplicate invoices, which are more common than most finance teams expect
6. Invoices for things with no purchase order at all

A rollout that does not frighten the finance team

Run in shadow mode for a month: the system processes everything, posts nothing, and a person compares. You will find your real accuracy rate rather than the brochure one, and finance gets to trust it before it touches the ledger.

Then go live by supplier, starting with the ten highest-volume ones, whose layouts are stable and whose invoices are most of your volume. Long tail last.

Frequently asked questions