AI for reading supplier invoices, and what still needs a person

6 days ago
3 min read

Updated: 5 days ago

Introduction

Invoice processing is the purest example of work that should not be done by a person. A document arrives, someone reads a supplier name, a date, a reference, a set of line items and a total, and types them into a system. It requires no judgement in the overwhelming majority of cases, it is done hundreds of times a month, and it is the task most likely to be behind when the month ends.

Automating the reading is now reliable enough for ordinary business documents, and the return is immediate. What determines whether the project succeeds is not the extraction accuracy but what happens to the small proportion it cannot handle, because that is where the time and the errors concentrate.

1. AI for reading supplier invoices is judged on the exceptions

Not on the average accuracy.

If ninety-five per cent are processed automatically, the remaining five per cent contain all the work. A clear queue, an owner and a target for those is what makes the difference between a saving and a mess.

2. Count your current volume and time first

The baseline for the business case.

Invoices per month, minutes each, and the hours consumed. Most owners underestimate this substantially because it is spread across several people and several days.

3. Match against purchase orders and receipts where you have them

The real control.

Extraction only replaces typing. Matching the invoice against what was ordered and what was received is what catches wrong prices, wrong quantities and things you never received, and that is where the money is.

4. Check line-level prices against your agreed rates

The recurring saving.

Invoiced unit prices against your price list, every line, every month. Discrepancies are common, small, and almost never queried because nobody checks. This alone frequently justifies the whole exercise.

5. Detect duplicates properly

Harder than it looks.

The same invoice arriving twice with a different reference, or a statement and an invoice for the same charge. Matching on supplier, amount, date proximity and description catches what a reference-number check misses.

6. Keep the approval step human

Extraction is not authorisation.

Someone still has to confirm the goods or service were received and that the charge is correct. Automating the reading and automating the approval are different decisions, and conflating them is how businesses pay for things nobody ordered.

7. Watch for the errors that pass silently

The characteristic failure.

A misread digit produces a plausible number that no rule rejects. Tolerance checks, running totals against supplier history and a periodic manual audit of a sample are what catch these.

8. Decide where the documents live and for how long

Both a practical and a compliance question.

Retention periods for accounting records are set by law and vary by jurisdiction, and tax authorities in some places have requirements about the form in which records are kept. Confirm what applies before discarding paper.

9. Measure the outcome as processing cost per invoice

The number that shows whether it worked.

Total time and cost divided by volume, before and after. Report it monthly alongside the exception rate, because a low cost per invoice with a rising exception backlog is not an improvement.

Watch the early payment discounts you were previously missing. Faster processing frequently unlocks terms that were always available and never achievable at the old speed, and that is sometimes the largest single benefit.

Conclusion

Design the exception handling first, because that is where all the remaining work sits.

Measure your current volume and hours to establish the business case, match invoices against orders and receipts rather than only extracting data, check every line price against your agreed rates, detect duplicates on amount and date rather than reference alone, keep approval as a human decision separate from extraction, put tolerance checks and a sample audit in place to catch plausible misreads, confirm the retention rules that apply to you, and report cost per invoice alongside the exception rate.