When invoice automation actually pays off
There is a threshold below which automating invoices costs more than it saves. Here is how to find yours, without buying an enterprise tool to discover it.
The three numbers that decide it
Ignore the vendor slide decks. Three numbers decide whether automating invoice entry pays:
- Volume — how many invoices a month, and how many lines each.
- Variability — how many different suppliers and layouts.
- Consequence — what a wrong number costs you (a supplier phone call, or a tax filing amendment).
Low volume plus high variability is the case where automation loses: ten invoices a month from ten wildly different suppliers takes longer to fix than to type.
The threshold, concretely
From working through small-business cases, the pattern looks like this:
- Under 20 invoices a month: type them. Automation overhead is not worth it.
- 20–100: this is the sweet spot for a converter. Enough repetition to make a template worth building, few enough that corrections are quick.
- 100+: now the question is not a tool, it's a process — approval flows, three-way matching, ERP integration. Different problem.
What "automation" should mean at the 20–100 mark
Not a dashboard. Not an integration. It should mean: the numbers arrive in a table, you fix the two that are wrong, you export. If the tool adds a workflow you did not ask for, it is costing you time.
The test is simple: does it save more minutes than it takes to learn? A tool that needs a 40-minute onboarding video has to save 40 minutes before it breaks even — which at 40 invoices a month is about two months.
Accuracy expectations, set honestly
Typed PDFs (exported from an accounting system) contain a text layer. Reading that is near-perfect — errors essentially only come from your parsing logic. Photographs and scanned PDFs go through OCR, where 90–97% character accuracy is normal, and the failure mode is predictable: faint print, skew, stamps over the total.
This is why every serious workflow ends with a human checking the totals. Anyone promising 100% is selling you a cleanup task.
The hidden cost of uploading
If the tool uploads invoices, you have taken on a data-processing obligation. In the UK that is a GDPR conversation; for accountants it is also a professional-ethics one. Worth asking: where is it stored, for how long, is it used for training, and can I delete it?
If the answer is vague, consider processing locally instead. A browser OCR engine is not as good as a paid cloud API, but "the file never left the laptop" is a sentence you can defend.
Common questions
How accurate is OCR on invoices?
For printed text, roughly 90–97% character accuracy on decent scans, higher on clean digital PDFs. Amounts and dates are usually right; supplier names and addresses are where mistakes cluster.
Should I OCR or read the PDF text layer?
Always try the text layer first. If a PDF was generated by an accounting system it already contains the text, so reading it is exact and instant. Only fall back to OCR for photos and scans.
Do I need an API?
Only at volume. Below about 100 invoices a month, a manual step of fixing two fields is cheaper than maintaining an integration.