Food supplier invoice OCR: how Écotable measures the footprint of restaurants

Écotable helps the restaurant industry through its ecological transition, with a label awarded to the most responsible establishments. Awarding it means auditing a full month of supplier invoices per restaurant, then tying every purchase line back to the reference environmental database. Since 2026, Koncile reads those invoices through an API and matches them against the Agribalyse reference base.

We are a company that supports players in the restaurant industry through their ecological transition. There is an invoice audit side, which is the part Koncile is involved in, and a labelling side for the most responsible establishments.
^{Lucas Biguet-Mermet, Tech Lead, Écotable}

Invoice auditing in the service of a label

Founded in 2018, Écotable runs an ecological label for the restaurant industry, backed by impact measurement tools and sector support. The label is only worth something if it rests on verifiable facts, and those facts sit in one specific place: the restaurant’s supplier invoices.

The protocol is easy to state and heavy to execute. Each candidate restaurant hands over a full month of purchase invoices. Every product line has to be extracted with its quantity and unit, then tied back to Agribalyse, the French environmental agency’s reference base carrying impact indicators for food products. That mapping is what turns a shopping list into a measured footprint.

The difficulty is the volume of data. A restaurant serving several dozen or several hundred covers a day can have dozens of invoices to process. Doing that by hand is fairly time-consuming.
^{Lucas Biguet-Mermet, Tech Lead, Écotable}

Across the platform that means several hundred invoices a week, and well over a thousand in some months, for a total of more than six and a half thousand documents processed since launch.

From market gardeners’ handwritten notes to wholesale invoices

What makes the restaurant industry distinctive is that its suppliers range from national food distributors to the grower delivering direct. The paperwork spreads just as widely.

We get invoices from small market gardeners that can be handwritten, and that are not always very legible, even for a human. Koncile copes rather well with those difficult cases.
^{Lucas Biguet-Mermet, Tech Lead, Écotable}

That heterogeneity is exactly what made Écotable switch. The company previously used a classic text recognition tool, which returned what was written on the invoice without understanding it. Koncile’s invoice OCR rests on semantic reading: the model interprets the document, which lets it handle a handwritten delivery note as readily as a structured multi-page invoice, with no separate template per supplier.

Normalising during extraction: kilograms and categories

The clearest gain is not in the reading itself, it is in what the engine does while reading.

With Koncile we can derive values from the fields written on the invoice. If the quantity is not in kilograms, we can ask for every quantity to be converted to kilograms so we have a single unit. We can do category detection: is the product detected a fruit or vegetable, is it a drink, is it something else. Beyond the extraction itself, that is what saves us time on processing the data afterwards.
^{Lucas Biguet-Mermet, Tech Lead, Écotable}

A case of twelve bottles, a crate billed by the unit, a weight in grams or in pounds: everything comes out in a common unit. And automatic categorisation files each line into Écotable’s own taxonomy at the moment of extraction. Those rules are written in plain language inside the model, and the teams refine them as edge cases come back from operations.

Tying every line to the Agribalyse base

That leaves the operation that gives meaning to everything else: linking “unsalted butter block” or “vine tomatoes class 1” to the matching Agribalyse entry, among roughly two and a half thousand products. No product code travels on the invoices, and every supplier writes its wording its own way. That mapping used to be done by hand, line by line, after extraction.

This is what Data Table Match does. Écotable loads the Agribalyse reference into the platform as a two-column table: the product wording on one side, its unique identifier on the other. An agent is attached to the extraction model and points at the description field of each invoice line. On every run it finds the closest entry and returns the identifier in an extra column, delivered straight into the data flow.

Matching runs in semantic mode. The base is vectorised and correspondence works on meaning rather than words, which finds the right product even when the supplier’s wording shares no term with the reference. Difficult matches can be steered with instructions written in plain language, added as use reveals them. The mechanism handles far larger references without difficulty, up to tens of thousands of entries.

This is exactly what we need for our use case. We were handling the Agribalyse mapping in our own tool after the import. Moving that step into Koncile takes one step out of the chain.
^{Lucas Biguet-Mermet, Tech Lead, Écotable}

An API integration, and an in-house review interface

The technical setup is entirely driven by Écotable, and Koncile stays invisible to the restaurateur.

We built an API integration. As soon as invoices are uploaded by our clients onto our platform, we send them to Koncile directly through the API. As soon as extraction has happened, the documents come back into our platform, and we built a review interface on our side. That lets us do everything within our own platform.
^{Lucas Biguet-Mermet, Tech Lead, Écotable}

The OCR API was wired up in one to two weeks of actual development, with separate test and production environments and versioned models, so extraction rules can evolve without touching production. What mattered to the operations teams sat elsewhere: previously they had to log into an external tool to check extractions, which added a step to every audit. It all happens inside Écotable’s own interface now.

What it changes for the teams

Human checking stays systematic, deliberately.

There is always human arbitration, we review every invoice. But successful extraction rates are very high, we have almost no corrections to make. We keep that check to guarantee data quality, but correcting by hand is fairly rare. Our operations teams can spend more time analysing the data than correcting it.
^{Lucas Biguet-Mermet, Tech Lead, Écotable}

A confidence score accompanies every extracted value and flags the lines worth a look, which concentrates checking where it earns its place. The shift is clear: the auditor no longer rekeys, they analyse. And on the economics, moving to Koncile cut the cost of extraction threefold compared with the previous solution.

What comes next: checking that audited documents are genuine

One workstream follows naturally, and it touches the value of the label itself. An audit that decides a certification mechanically creates an interest in presenting a more favourable file than reality: an invoice retouched to substitute one product for another, a document fabricated in a local supplier’s name, a file reused from one period to the next. Across several hundred documents a week submitted by the establishments themselves, checking everything by eye is not realistic.

Koncile’s document fraud detection answers that risk by analysing the file as much as its content: forensic examination of the PDF metadata and the last modifying software, visual integrity analysis around a quantity or a wording, consistency checks on whether amounts and dates are plausible. Each anomaly feeds a score with its justifications, returned alongside the extracted fields and the Agribalyse matches, in the same call. For a body that awards a label, the stake is direct: the credibility of the certification rests on the credibility of the documents behind it.

Automate the reading of your supplier invoices

Purchase invoices, delivery notes, handwritten documents: if your teams rekey product lines and map them to a reference by hand, send us your real documents. We configure your fields, your normalisation rules and your lookup table, and you judge on your own paperwork.

Book a demo