Re-read a complete export file, document by document, and flag every inconsistency before the container leaves.
An export file brings together about ten documents issued by different parties: the customer's purchase order, invoices, packing lists, certificates of origin, analysis, health or phytosanitary certificates, the bill of lading. One quantity that differs between invoice and packing list, one mistyped container number, one missing certificate, and the goods are held at the port or payment is delayed. This tool reads every document (digital or scanned), extracts the key data while citing exactly where it read it, then runs a set of cross-checks built around one simple rule: the purchase order is the reference. The output is an HTML report with a built-in document viewer to verify each alert, plus correction request e-mails drafted per issuer, in French and English, that the exporter reviews and sends personally. The project is at pilot stage: it was delivered as source code to an exporter, and day-to-day use by the client is not yet established.
For the exporter's end customer, the expected effect is goods that arrive with consistent documents the first time. For partners (forwarders, laboratories, banks), it means precise correction requests that cite the page and the expected value, instead of a string of approximate e-mails. For the export team, it is a systematic review that nobody has to carry in their head any more.
The export team drops the file's PDFs into a folder. The tool extracts quantities, weights, prices, batch references and container numbers from each document, compares them with the purchase order and produces the report. A net-weight gap between the packing list and the certificate of analysis shows up as an alert, with both pages side by side. The manager decides whether to correct it or treat it as an accepted tolerance.
A health certificate carries a batch number that does not match the invoice. The tool drafts an English e-mail to the issuing laboratory, citing the document, the line concerned and the expected value. The exporter reviews it, adjusts the tone if needed and sends it from their own mailbox. The tool sends nothing.
Before handing a file to its bank or submitting it to an export support scheme, the company runs the full set of documents through the checker. The required-document list and the consistency between documents are checked one last time; the report serves as evidence of internal review.
The report explains each alert in plain language (which document, which rule, which value was expected). A recent hire uses it as an annotated checklist and learns the rules of the export file by seeing them applied to real cases.
This product applies to the following business functions. Click to discover how we adapt it to each.
These products combine with this one for complete programs.
AI-augmented search on your document heritage — answers grounded in your documents, with sourced citations.
Keep humans in the decision on critical cases, at scale.
Run multiple specialized AI agents together, on real business task chains — without breakage and without rogue agents.
Two internal measurements: on a reference file, all 45 expected fields were extracted correctly (recorded 16 September 2026); on a bench of anomalies deliberately injected into files, 34 out of 35 were detected and no false alert was raised (recorded 7 October 2026). Known limit: the reference file was built by us and has not yet been validated by the pilot client.
At pilot stage. The code was delivered to an exporter, but day-to-day use by the client is not established at this date. We therefore present it as a pilot, not as a product in service. Deployment at your company starts with a trial on a few of your real files; that trial is what then lets the schedule be settled at scoping.
Yes. The text of every document is extracted locally for native PDFs and through an online OCR service for scans; in both cases that text is then sent to a Mistral language model hosted in the European Union for data extraction. If that constraint is a blocker, a fully local variant can be studied at scoping, on the same principle as our contract analysis on a local server.
That is the intent. The required-document list, accepted tolerances (weights, quantities, rounding differences) and partial-shipment handling are parameters, not hard-coded logic. Every rule added or changed is covered by an automated test, so tomorrow's adaptation does not break today's control.
Neither. It drafts the correction e-mails, it does not send them; it flags anomalies, it blocks nothing. The decision to ship, correct or accept a discrepancy stays with the exporter. An automatic control acting on its own in a logistics chain would create more risk than it removes.
Free initial scoping, 30 minutes to 2 hours, with a solution lead adapted to your sector.