The €825,000 you saved before lunch

9:47 AM. Your phone buzzes.

“Forges Martellière just became insolvent. Our exposure is €1M+. Executive call at noon.”

Your stomach drops. Because you know what €1M+ actually means.

It means seven subsidiaries across France, the UK, the US, Germany, Italy, Spain, and the Netherlands. It means analysing 1,004 documents scattered across local SharePoints, legal vaults, procurement folders that nobody has opened in eighteen months. It means perusing over contracts in French, English, German, Italian, Spanish, Dutch, six languages, three legal systems and two currencies.

It means picking up the phone.

Calling François in Paris. Then Klaus in Stuttgart, who is in a meeting. Then leaving a voicemail for Sophie in Madrid. Then Maria in Milan, who needs to ask her assistant. Then the UK office points you to legal, who then points you back to procurement.

Each of them opens a folder. Each of them reads. Each of them calls back "I think we have something, let me check with..."

Great!

It is 10:18. You have eleven people on email. Three of them have replied. None of the replies agree.

The call is in 102 minutes.

You stop. You open LightOn. You type one sentence. In French. Because that's how you think.

"Forges Martellière est en redressement judiciaire. Quels BDC en cours ? Quelle est notre exposition ?"

No filters. No country tags. No language hints. No document-type scoping. Just the question, the way you'd ask François.

You hit enter.

LightOn fans out across all seven subsidiaries at once. It reads French contracts in France, English POs in the UK, German master agreements in Stuttgart. It runs searches. Two return and nothing is itself the answer you are searching: the UK has zero open commitments. The US has zero open commitments. Silence, confirmed.

Six minutes.

• One open PO. Véracier Défense, France. €175,000.
• Section 9 of the Master Supply Agreement, identical across all three jurisdictions: immediate termination on insolvency. Cited. Sourced.

You close eleven email threads. You tell François to stand down.

Noon. Executive call.

Legal opens with the €1M fallout. The CFO braces for impact.

You speak second.

"The real number is one-seventy-five. The million is the panic estimate from the email. We have immediate termination rights in all three jurisdictions. Here are the document IDs."

The room recalibrates around your number.

Without LightOn, you would have spent the morning calling eleven people, reading three contradictory replies, and walked into the call echoing the email because you had nothing harder.

With a DIY RAG, you would have confidently repeated "approximately €1M" because the panic email said so five times.

You didn't because you knew an email is not a contract. And you had the receipts.

The story sounds simple in hindsight: one question, six minutes, one defensible number. But the important part is not that LightOn “found a document.” It is that the system refused to stop at the first plausible answer. It separated panic from proof, searched across entities, confirmed absences, and rebuilt the exposure from source authority rather than semantic proximity.

Run the test yourself

The Forges Martellière scenario is part of EDiTh, LightOn’s open enterprise dataset built around Véracier Industries: a fully synthetic industrial group with 1,004 documents across seven subsidiaries, six languages, scanned PDFs, bilingual contracts, internal emails, and procurement artifacts.

It was built for one reason: so you can test whether an AI system can handle enterprise evidence before you trust it with your own.

Run the same question:

"Forges Martellière est en redressement judiciaire. Quels BDC en cours ? Quelle est notre exposition ?"

See whether your stack repeats the loudest number in the corpus, or finds the number you could actually take to the CFO.

Start with EDiTh. Then run it on your own documents.

Get access to LightOn Console to run the scenario yourself.

Want to understand how the corpus was built, how the retrieval worked, and why this answer is hard to get right? Read the EDiTh launch post.