Invoice processing is the most requested first automation, for good reason: it is high volume, entirely manual, universally disliked, and the cost is easy to measure. It is also the one most likely to be built badly, because the failure mode — a wrong number reaching your accounting system — is silent and expensive.
Here is how it should be built. If you would rather have it built for you, the pricing is on our workflow automation page.
What the manual process actually costs
Count invoices per month, multiply by minutes each takes end to end — receiving, reading, coding, matching, chasing approval, entering, filing — and multiply by the loaded hourly cost of whoever does it. Most small finance teams are surprised by the total.
The ROI calculator does this and subtracts the running costs, so you get a payback figure rather than a savings figure.
The five stages
1. Capture
Invoices arrive by email attachment, supplier portal, post and occasionally fax. Route everything to one place — usually a dedicated inbox that the automation watches. Paper gets scanned into the same place.
The goal is one entry point. Automations that watch four different sources are four times the maintenance.
2. Extraction
Reading supplier name, invoice number, dates, line items, tax and total from a document that looks different for every supplier. This is where AI genuinely outperforms older template-based OCR, because it copes with layouts it has not seen before.
The critical design decision: every extracted field carries a confidence level, and anything below your threshold goes to a human rather than through. Do not skip this.
3. Matching
Against the purchase order and the goods receipt, if you use them. Quantities and prices should reconcile. Mismatches are exceptions, not errors to be resolved automatically — the system flags, a person decides.
4. Approval
Routed by rules you set — amount thresholds, department, cost centre. Approvers get a link, approve or reject with a comment, and the system chases the ones sitting idle.
Automating the chase is most of the value here. The delay in accounts payable is rarely the data entry; it is waiting for someone to look at it.
5. Posting and filing
Into your accounting system with the correct coding, and the document filed where it can be found in an audit.
The three mistakes that cost money
Trusting extraction without confidence thresholds. Extraction is good, not perfect. A system that posts whatever it read is a system that will eventually pay $12,400 against an invoice for $1,240. Route low-confidence fields to a person — the volume is small and it is what makes the whole thing safe.
Automating approval as well as routing. Route automatically, approve manually. The moment money leaves without a human deciding, you have built a fraud vector. Invoice fraud usually works by looking routine.
Ignoring the exception queue. Every automation has one, and an unwatched queue silently accumulates unpaid invoices and unhappy suppliers. Someone must own it, and it must be visible daily.
How to sequence the rollout
- Run extraction in shadow mode first — it reads, a person still keys in. Compare for two weeks and measure the accuracy on your real invoices.
- Turn on posting for your highest-volume, lowest-risk supplier only.
- Widen supplier by supplier, watching the exception rate.
- Add approval routing once posting is stable.
- Keep approval thresholds conservative for the first quarter, then loosen if the numbers justify it.
The shadow-mode step is the one people skip and the one that tells you whether the system is safe to trust. It costs two weeks and it is worth it.
What good looks like after six months
- The large majority of invoices flowing through without human touch.
- A small, visible exception queue that someone clears daily.
- Approval cycle measured in hours rather than days.
- Every invoice findable in seconds during an audit.
- A monthly report showing volume, exception rate and time saved.
Frequently asked questions
How accurate is invoice extraction?
Good enough to be useful, not good enough to go unchecked — and anyone quoting a single number across all invoice types is guessing. Build a test set from your real invoices and measure before go-live. You should see the actual figure on your own documents, and you should keep the test set to re-run later.
What about invoices in other currencies or languages?
Both are handled well by modern extraction. Currency conversion should use a rate source you control and be recorded on the transaction, not silently applied.
Will it work with our accounting software?
QuickBooks, Xero, NetSuite, Sage and most mid-market systems have APIs. Older systems usually accept a structured import file. Tell us what you run and we will say plainly whether it is straightforward.
How long does it take to build?
About two weeks for a single-entity setup with one accounting system, longer where there are multiple entities, purchase-order matching or unusual approval rules. The shadow-mode period is on top of that and is worth budgeting for.
Want this built properly? We map the process free first and tell you honestly whether the volume justifies it. See workflow automation pricing or message us on WhatsApp.
