How a document colleague works.

What you hand over

A copy. Not a migration, not a new system to move into, not a year of preparation. A daily copy of the archives you already keep: the data room, the mail archive, the accounting export, the shared drive, the scans nobody has opened since they were filed.

Nothing is read before you have said so in writing, in a data-processing agreement with Axon Trade ApS, the Danish company behind Neural Logic. You say which archives are in scope, and that is the boundary.

Conversion: files become an addressable archive

Every document is read once, properly. A scan becomes text, a contract becomes clauses, a mail thread becomes dated messages, a protocol becomes decisions. Each piece keeps a link back to the page it came from, so any sentence can be traced to its original.

The result is an archive you own: checked, addressable and usable with any language model, ours or the one you choose in three years. That is deliberate. A client who cannot leave is not a client, only a captive.

The register: what is there, and what is not

While the documents are read, a register is built: which contract carries which clause, which protocol is signed, which invoice belongs to which case, which document was expected and never arrived. The register is a record, not a memory.

This is what makes “not found” a real answer. When the colleague says a termination clause is missing, that comes from the register, which knows every document it read and every place the clause could have been. It is not a model failing to recall something.

Reading a case

You ask about a case, a property, a client, a contract or a project. The colleague assembles the complete file in date order across every archive in scope, quotes the passages that decide the question word for word with a link to the original, and states plainly what it looked for and did not find.

What comes back is a file, not a chat answer: the documents, the order they happened in, the decisive words, and the gaps. A person reads it and decides.

Every quotation is checked

Before a quotation reaches you it is compared, word for word, with the source document. If it does not match, it is held back and you are told. A colleague that quotes a contract almost correctly is worse than no colleague at all, because the mistake looks exactly like the truth.

Two roles, one boundary

The colleague who reads your documents runs on a European model, Mistral, processed in Europe under a European contract. Documents do not leave that boundary.

Behind her stands a supervisor that never sees a document or a name. It reads only the shape of the work: what kind of question was asked, how many documents were read, which word was missing, where the reading went wrong. Because nothing of yours reaches it, it is chosen for thinking rather than for location: today that is Claude, from Anthropic, in our view the best model for this kind of task. It improves the colleague a little every night, without any access to your data.

What it is not

• Not a chatbot on your intranet. It answers from a checked register built while reading your documents, not from a model's recollection of them.
• Not a decision-maker. It lays out facts. A person weighs them and acts.
• Not a judge of people. Files on buildings, products, contracts and cases. No scores, no rankings, no predictions about people.
• Not a writer into your systems. Writing back is off by default and switched on only by invitation, where you ask for it and the rules allow it.

Where this stands today

Frida, the document colleague, is in preparation, with the first client engagement being prepared now. Vera, the supervisor, is in preparation. Pia, a catalogue colleague built on the same method, runs every day on a web shop of our own, reading a supplier's feed and keeping the shop true to it.