Extracting a date, an amount or a VAT rate from a PDF: AI can do that today. In a demo, it all looks simple. You drop in a receipt, the data lands in the accounting software.
The real subject is elsewhere. An AI that fails visibly poses few problems, because someone notices. The difficulty is the plausible error, the one that looks exactly like a good answer.
Reading an invoice has become accessible
A few years ago, automating invoice entry meant configuring each supplier. Every new format required a setting. Today's AI models read a document they have never seen and extract the useful fields.
The gain is real. Manual entry of accounting documents is repetitive, unrewarding and prone to typos. Handing it to a machine frees up time for analysis and supplier follow-up.
But a demo shows the favorable case: a clean invoice, a single rate, a known supplier. It says nothing about the day the document falls outside the ordinary.
The exceptions are the real difficulty
In an accounting department, special cases are not rare. They are part of the daily routine:
- a down payment, which must not be treated as a settled invoice;
- two VAT rates on the same invoice, to be allocated correctly;
- a credit note read as an invoice, turning a refund into an expense;
- a document sent twice, by email and then by post, at risk of being paid twice.
On these cases, AI does not always signal its unease. It can produce a perfectly plausible answer, and a wrong one. The amount has the right format, the date is consistent, nothing catches the eye.
This behavior stems from the very mechanism of these models. They produce the most probable answer given the document. They do not redo the arithmetic and do not check that the net and the VAT add up to the gross. A fluent answer is not a verified answer.
The real questions then become organizational. Who checks the amounts? Who spots the duplicates? And who takes over when something is off?
Cost is steered document by document
The second blind spot is economic. Running every receipt through a powerful model has a price. Some pipelines even call the model several times on the same document, to read, classify, then check.
On a few dozen invoices, the expense goes unnoticed. On thousands a month, it becomes a budget line in its own right. And the tool's overall bill does not say where the money goes.
The useful indicator is the cost per document. It shows which kinds of documents cost a lot, and makes it possible to ask whether they really deserve the most powerful model.
Electronic invoicing does part of the work
Since September 1, 2026, every company in France must be able to receive electronic invoices, according to the reform's timetable. Large companies and mid-caps already issue them. Smaller companies will follow on September 1, 2027.
These invoices arrive in structured formats (Factur-X, UBL or CII). The amounts, the rates and the references are already machine-readable. For them, asking an AI to "read" the document makes little sense.
AI keeps its usefulness for the rest: foreign suppliers, till receipts, expense claims, PDF invoices still received during the transition. The reform does not remove the subject. It invites you to reserve AI for the documents that really need it.
A control circuit in four tiers
The solution is not to hand everything to AI, nor to give up on it. It consists in organizing a circuit where each document gets the treatment it deserves, no more and no less.
- Simple tools for simple documents. A structured invoice, or a recurring supplier with a stable format, is handled by rules. It is fast, predictable and cheap.
- Automated checks on every document. Net plus VAT equals gross, authorized rates, known supplier, invoice number already seen for that supplier. These checks cost next to nothing and block any arithmetic inconsistency.
- The most powerful AI for complex cases. An unusual document, a layout never seen before, a failed check: that is where the most capable model brings its value.
- Human validation for doubtful cases. When a check fails or the model reports low confidence, the document goes to a person. They decide, and their decision is recorded.
This circuit changes the nature of accounting work. You no longer re-enter every document. You handle the exceptions, with the context needed to decide quickly.
Traceability keeps every decision legible
Automating must not mean losing track of decisions. For each document, you need to be able to answer simply: who handled it, with which tool, and who validated it in case of doubt.
That trace is useful day to day, when a supplier disputes a payment. It is also useful during an audit, to explain how an amount entered the accounts. An error spotted late can then be corrected at the source, without an investigation.
Finally, it touches on control of your data. Your invoices describe your suppliers, your prices and your volumes. Knowing where they are processed, by which model, and being able to change that without rebuilding everything is part of the same subject.
Two figures say where you stand
At the outset, the question seemed technical: can AI read an invoice? The answer is yes, and that is not the question that counts. The right question is twofold: how many errors go unseen, and how much does a processed document cost?
A tool unable to answer either one is not yet a production system. Take the last twenty invoices received and count those that fall outside the ordinary: down payment, credit note, several rates, duplicate. That number tells you where the reliability of your automation will be decided.
A reliable AI is not one that never doubts. It is one that knows when to ask for a check, and whose cost you keep in hand.