Contents7
An invoice arrives by email. Someone opens the attachment, re-keys the amount and the number into the accounting software, looks for the matching purchase order, checks that the delivery actually happened, then passes it on for approval. Multiplied by hundreds of invoices a month, this chain takes up whole days, and it is one of the easiest processes to automate properly. This guide from our section on process automation and business bots breaks it into three steps, and says for each what belongs to a rule, to AI or to a person.
In short
A supplier invoice is processed in three steps: read, match, approve.
Since 1 September 2026, every business in France must be able to receive electronic invoices, whose data can be read without AI.
AI is only needed to read unstructured documents: receipts, PDFs, photos.
Matching against the order and the delivery is a matter of rules and tolerances.
A change of bank details is never approved automatically.
The three steps of a supplier invoice
An invoice’s path, from receipt to payment
- Read
Extract supplier, number, date, amounts and VAT.
- Match
Compare with the order and the delivery.
- Approve
A person signs off what falls outside the rule.
- Pay and file
The entry and the document, in the right place.
Each step is automated differently, and confusing them is what produces disappointing projects: an AI asked to do everything, when two steps out of three are settled by written rules.
Reading: e-invoicing changes the picture
For a long time, reading an invoice meant reading a PDF whose layout changed from one supplier to the next, hence character recognition, then language models. The French e-invoicing reform changes that starting point. According to the French tax administration’s page:
- 1 September 2026
Every business must be able to receive electronic invoices; large and mid-sized companies must also issue them.
- 1 September 2027
Small and micro-businesses must in turn issue electronic invoices.
An electronic invoice, in the sense of the reform, is not a PDF sent by email: it is a structured data file (UBL, CII, or a mixed format that attaches the data to an image), sent through an approved platform. Its amounts, VAT and references are read without AI, field by field. For the invoices concerned, the most expensive step to automate becomes the simplest.
There will still be documents to read, though: receipts, expense claims, some invoices from foreign suppliers, older documents. There, a language model helps, provided it never guesses. A field missing from the document must come out as “missing”, not be filled in with the most likely value.
From the field
Reading reveals what is missing
On an online plush-toy store, extracting attributes from the free text of 481 product pages showed that only 41% gave a readable size. The lesson applies to invoices: before automating the reading, measure what your documents really contain, and plan the path for incomplete ones.
Matching: a matter of rules
Matching compares the invoice with what you actually ordered and received. People speak of two-way matching (invoice and order) or three-way matching (invoice, order and goods receipt). None of these comparisons needs AI:
- does the amount match the order, within the tolerance you set (rounding, an agreed shipping difference)?
- were the quantities invoiced actually delivered?
- has the invoice already been received, under the same number or with the same amount on the same date?
- is the supplier known, and are their bank details the ones on record?
An invoice that passes all these checks can go on alone. An invoice that fails one goes to a person, with the reason written down. This is exactly the split described in our guide to automation without AI, when a rule is enough: the same input always gives the same result, and there is nothing to hallucinate.
Key takeaway
Matching is only as good as your order and receipt data. If purchase orders are not entered, there is nothing to match: that is often the first project, before the automation itself.
If your accounting software or ERP offers no API to write into, there are other ways in, which we compare in six ways to automate software without an API.
Approving: what stays with a person
Automation sorts; a person approves. Three cases should never go through without a human look: a new supplier, a gap beyond the tolerance, and a change of bank details.
That last case is the entry point for fake-supplier fraud. La Banque Postale, a French bank, pointed out in January 2025 that more and more fraudulent payment requests rely on a genuine invoice whose bank details alone have been changed. It recommends contacting the supplier on their usual phone number, and splitting the entry and the approval of bank details between different staff members.
Watch out
A rule to write into the workflow
Bank details that differ from those on record block the invoice and alert a person, who calls the supplier back on the number already known. The automation never updates bank details from an invoice or an email.
The accountant, in-house or at a firm, stays at the centre: they set the posting rules, approve special cases, check VAT and close the books. What automation takes off their hands is re-keying and chasing documents, not judgement.
What automation changes, and what it does not promise
The gains show first in regularity: no more invoices forgotten in a mailbox, a duplicate spotted before payment rather than after, a history of every decision, suppliers paid on time. What it does not do: repair missing order data, or decide on someone’s behalf in a dispute with a supplier.
From the field
We run this split of roles on our own expenses. Our bookkeeping entry agent, in production since July 2026, reads the photo of a receipt sent on Telegram and extracts the supplier, the amount, the VAT, the date and the category; checks look for duplicates, amount consistency and a confidence threshold before anything is written, and a person approves the entry with one button. It is one of the tools we run ourselves.
Where to start
- Count your documents.
How many invoices a month, in which formats, from how many suppliers.
- Check your orders.
Without entered purchase orders, start there.
- Write the rules.
Tolerances, duplicates, thresholds, and who approves what.
- Automate reading last.
Electronic invoices first, PDFs and receipts next.
Our n8n automation, with or without AI starts from the task as your team does it today, and shows the cost of each run. The same flow can feed a weekly report without copy-and-paste of pending invoices and gaps.
To go further in the section: n8n, Make or Zapier for a small business, monitoring n8n workflows, the n8n licence, customer support sorted by AI, WhatsApp and website bots, employee onboarding, AI voice agents on the phone and AI and young people’s first jobs.
Frequently asked questions
Can AI replace the accountant for supplier invoices?
No. It can read an unstructured document and propose an entry, but posting rules, VAT, special cases and closing remain the accountant’s job. Automation takes re-keying and document hunting off their hands, which leaves more time for control and advice.
Do you need AI to read an electronic invoice?
No. An electronic invoice in the sense of the reform is a structured data file: amounts, VAT and references are read field by field, by a simple rule. AI is only justified for documents that remain PDFs or photos, such as receipts or some foreign invoices.
What is three-way matching?
It is the comparison of three documents before paying: the purchase order (what you asked for), the goods receipt (what you received) and the invoice (what you are asked to pay). If all three agree, within tolerance, the invoice can be paid; otherwise it goes to a person with the gap found.
How do you avoid paying a fake invoice?
By blocking any invoice whose bank details differ from those on record, and verifying the change with a call to the supplier’s usual number. Also split the entry and the approval of bank details between two people. A well-written automation applies these rules to every invoice, without exception.
Sources
- The French tax administration’s page (new tab)impots.gouv.fr
- La Banque Postale, a French bank, pointed out in January 2025 (new tab)labanquepostale.fr
La Banque Postale, a French bank, pointed out in January 2025
Read on labanquepostale.fr(new tab)Jim · founder of Brain Plus AI
Also in automation and bots
Choose and price9 min
Automation without AI: when a rule is enough
Not every automation needs AI. A fixed rule costs less and always gives the same result: here is how to spot the steps where AI really pays off, and those where it only adds cost.
Practical guide7 min
Automating software without an API: six ways in
Your business software has no API, or keeps it for a more expensive plan? There is usually still a door for automating it: here are the six, from the most solid to the most fragile, with what each one costs in maintenance.
Practical guide6 min
Automating reports: the end of Monday copy-and-paste
The Monday report rebuilt by hand, copying figures from one tool to another, can almost always be automated without AI. What an automatic report should contain, where its figures come from, and the safeguards that stop a wrong report from going out on time.