Supplier invoice entry: what AI reads well, and badly

Supplier invoice entry: what can an AI read?

An AI reads well what the invoice states in plain print: supplier, number, date, amounts, VAT, due date. It reads badly what needs context: a discount agreed over the phone, a down payment, an ambiguous line, a new supplier, a damaged scan. It prepares the entry; a person decides the exceptions. That split is what makes automation workable in a small business, not the promise of perfect reading.

Supplier invoice entry is the classic job of whoever “does the books on top of their real job”: open a PDF, retype eight fields, pick an account, file the document. This article says where an AI is reliable, where it is not, and where to place human control. It complements the processes the studio automates: here we describe what you will actually see on your desk.

What the published guides say, and what is missing

We read the guides from Cegid, Lido and Yooz on invoice OCR and matching. They agree on the mechanics: receipt, extraction, matching against the order, validation, posting. Two points stand out.

• Accuracy figures come from vendors. Lido, for example, claims input errors falling from about 2% to under 0.2%, and 10 to 15% of invoices still needing manual handling after matching. These are commercial claims with no independent study. Cegid does not quantify accuracy and states that OCR does not replace human control.
• Nobody describes a ten-person company. The guides assume accounting teams and purchase orders for everything. In your business, half the invoices arrive without an order, by e-mail, sometimes as a photo.

Keep the order of magnitude, not the percentage: a large share of routine invoices processes itself, a minority goes back to a person. Your real ratio can only be known by testing on your own invoices.

What an AI reads well

Current engines no longer need a template per supplier: they read a layout they have never seen. They are reliable where the information is explicit and checkable.

• Header fields. Supplier name, invoice number, date, currency, due date.
• Totals. Net amount, VAT, amount due, with an arithmetic check available: net plus VAT must equal the total.
• Payment details. IBAN and reference, to compare with what you already hold for that supplier.
• Recurring invoices. Rent, subscriptions, energy: same supplier, same structure, month after month.
• Filing. Naming the file, storing it in the right place, spotting a duplicate by number and amount.

What an AI reads badly

An AI mostly fails when the invoice does not contain the answer. The amount is read correctly, but the right decision depends on a fact that is absent from the document.

An extraction error cascades: a misread amount later creates a matching gap or, worse, a wrong payment. The danger is not an AI that visibly fails; it is an AI that answers confidently on an ambiguous document. That is why control does not sit at the end of the line, it sits on the risky cases. We cover these failure causes in our article on AI automations that fail on real-world data.

Who decides: five validation rules

The principle: the AI prepares, a person validates whatever falls outside the norm, and the norm is written down in advance. These are the rules we usually set, to adapt to your business.

1. Arithmetic check. Net plus VAT equals total. If not, the invoice is set aside.
2. Known or unknown supplier. A supplier never seen before is always validated by a person. After that, it joins the list.
3. Amount threshold. Above a threshold you set, validation is mandatory even if everything is consistent.
4. Deviation from history. Amount far from that supplier’s average, changed IBAN, new due date: alert.
5. Duplicate. Same supplier, same number or same amount within a few days: blocked.

A changed IBAN deserves special attention: it is the classic pattern of bank-detail change fraud. The tool must never pay to an IBAN different from the known one until a person has confirmed it outside e-mail, for instance by phone.

Where the AI brings more than data entry

The most visible gain is retyping time. The least visible gain is consistency. A tired person on a Friday evening, at the tenth PDF, lets a duplicate or a swapped date slip through; a rule applied to every invoice does not get tired. The AI therefore also acts as a checker: it compares each invoice with the supplier’s history, spots the number already seen, flags the due date that changed.

This assumes one condition: that the rules are written down. An AI without rules proposes, but does not know when to stay quiet. The rules in the previous section are what turn a document reader into a flow you can rely on. They can be changed in minutes as you discover cases nobody had foreseen.

How to read your own results

Measure three things, not one. The share of invoices passed without intervention, the error rate among them (found by the weekly sample), and the average time spent on an exception. A high share of automatic invoices with a sample that reveals errors is a bad sign: the rules are too loose. A low share with few errors is a good start: the flow is cautious, and you will loosen it supplier by supplier.

A worked example, with its assumptions

Take a company that receives 200 supplier invoices a month. Assumptions, to replace with your figures: five minutes of entry and filing per invoice, about 17 hours a month. Suppose 75% of invoices pass the five rules untouched and 25% land on a person’s desk for two minutes of checking each. Human time falls to roughly 1 h 40, plus the sample check of invoices that went through alone. These proportions are a working assumption, not a result measured at a client: yours depends on document quality and on the share of regular suppliers.

The useful calculation is therefore not “how often is the AI right” but “how many exceptions arrive each week, and how many minutes does each take”. To size this time for your case, the formulas and prices are a starting point for the conversation.

How to start without taking risks

1. Take fifty real invoices from the last three months, good and bad, scans included.
2. Have the AI read them and compare, field by field, with the existing entry. Note the errors and their type.
3. Write the validation rules from those errors, not from a generic list.
4. Run in parallel for a month. The AI prepares, you still enter as before, and you compare.
5. Switch over supplier by supplier, starting with the regular ones.
6. Keep a sample check: one invoice in twenty, reread every week, even when all is well.

When this is not the right choice

• Fewer than twenty invoices a month. Manual entry takes an hour or two. Setup time will not be recovered.
• Mostly handwritten or poor-quality documents. Improve receipt first (ask suppliers for native PDFs).
• No identified owner. Without a person who receives the exceptions, they pile up and the flow loses the team’s trust.
• Your books are kept by a third party who requires their own process. Talk to them first: the flow must feed their tool, not duplicate it.

What we do not promise

We do not promise a reading rate. A figure quoted without your documents is worthless, and the ones vendors publish cannot be checked from outside. We promise a method: a test on your invoices, written rules, a sample check, and a flow that stops when it doubts. Questions of retention, VAT or e-invoicing are for your accountant or adviser, not for this article.