Contents
AI extracts the supplier, item numbers, quantities and amounts from delivery notes and unstructured invoices, then matches them against orders to flag discrepancies. With e-invoicing in France, it mostly serves delivery notes and documents outside the mandate. A first use case starts at $2,300 (€2,000)excl. VAT.
Delivery notes and supplier invoices are among the most retyped documents in any company: someone copies the quantities received, compares them with the order, and enters the invoice into accounting. AI can read these documents and extract the data, but France’s e-invoicing reform changes the picture for some of them. Here is what AI extraction does well, where it remains useful, and how I set it up.
What e-invoicing changes in France
In France, the rollout of mandatory business-to-business e-invoicing began on September 1, 2026. According to the French tax administration’s fact sheet updated in October 2026 (in French), all companies must be able to receive e-invoices through an approved platform since that date, and small and mid-sized companies must issue them by September 1, 2027 at the latest (checked on October 27, 2026; see also the overview page on impots.gouv.fr).
The practical consequence: invoices from your French suppliers will increasingly arrive as structured data that no longer needs extracting. This page is not a substitute for advice from your accountant on your obligations; the page on mandatory e-invoicing in France explains how to adapt your tools.
Where AI extraction is still useful
- Delivery notes, which are not invoices and often remain PDFs, scans or photographed signed sheets.
- Invoices from foreign suppliers, which generally do not go through the French platforms; your accountant can tell you how to handle them.
- The transition period, while some of your suppliers still send PDFs.
- Quotes, order acknowledgments and carrier slips, which contain useful but unstructured data.
- Three-way matching: purchase order, delivery note, invoice. Even when the invoice is structured, AI helps match its lines to those of a scanned delivery note.
Documents and checks
| Document | Data extracted | Matched against | Human check |
|---|---|---|---|
| Delivery note | Supplier, number, date, items, quantities delivered | Purchase order | Quantity or item discrepancies |
| PDF invoice (foreign, transition) | Supplier, number, date, lines, amounts, VAT | Order and delivery note | Price or amount discrepancies |
| Order acknowledgment | Items, confirmed quantities, lead time | Order sent | Changed lead times or quantities |
| Carrier slip | Shipment number, parcels, date | Planned shipment | Missing parcels |
The core value is not the extraction itself, but automatic matching and flagging of discrepancies: an incomplete delivery, a price different from the order, a substituted item. People no longer check everything, only what does not match.
What it costs
These amounts exclude VAT and come from my pricing.
| Situation | What is delivered | Indicative budget | Timeline |
|---|---|---|---|
| First use case on your current tools | Extraction for one document type, review screen, export | From $2,300 (€2,000)excl. VAT | 2 to 8 weeks |
| Module built into a business application | Extraction, order / delivery / invoice matching, accounting connection, dashboard | $6,800 to $22,400 (€6,000–€20,000)excl. VAT | 1 to 3 months |
In the estimator, built-in AI counts for $2,800 (€2,500)excl. VAT, a connection to another tool (accounting software, ERP) for $2,300 (€2,000)excl. VAT, the quotes, invoices and PDF block for $1,700 (€1,500)excl. VAT, and a dashboard for $1,700 (€1,500)excl. VAT. AI API usage comes on top, billed by the provider. Time savings are not promised upfront: they are measured on your documents during the prototype.
Safeguards
- A confidence score per field, with human review below a threshold set with you.
- The original document always available next to the extracted data.
- A change log: what the AI read, what was corrected, and by whom.
- Limited data sent out, only what is needed; see what data to share with AI.
- No accounting entries without review at the start. Syncing with your software is covered in accounting software integration.
When it is not the right solution
- Your invoices already arrive in a structured electronic format through your platform: use that data, there is nothing to extract.
- Your accounting software or receiving platform already offers document recognition and you are happy with it.
- Volume is low: a few delivery notes a week are quick to enter.
- Your main suppliers could use EDI or a portal: structured data at the source beats extraction.
Where to start
- Gather about fifty real documents of the most retyped type, usually delivery notes.
- Describe the current check: who compares what, in which tool, and which discrepancies cause problems.
- Run a one-week audit and a prototype on those documents before deciding on a rollout.
The full approach, with human review kept in place, is on the AI integration page.
Frequently asked questions
Does e-invoicing make AI extraction pointless?
For invoices received from French suppliers in a structured electronic format, extraction is no longer needed: the data already arrives machine-readable. AI remains useful for delivery notes, invoices from foreign suppliers, quotes, and during the period when not all of your suppliers issue e-invoices yet.
How reliable is the extraction?
It depends on document quality: a PDF generated by software reads very well, a crooked scan or a handwritten note much less so. That is why each field gets a confidence score and doubtful documents go through human review. Real-world reliability is measured on your documents during the prototype.
Can extracted data go straight into accounting?
Yes, through a connection to your accounting software or an import file, but only after review at first. Later, simple documents from a regular supplier can go through automatically, once measurements show the extraction is reliable.
Where are the documents stored?
In your application or your storage space, hosted according to your requirements. Only the content needed for extraction is sent to the AI API, with a provider whose business terms exclude training on your data.