Private AI is easy to claim and harder to show. This is the most common Euclid AI workflow, step by step, and what each step keeps inside the firm's control.

The argument for private AI in commodity trading is settled at the level of principle. Send your contracts, positions and P&L to an external model and you hand your commercial edge to infrastructure you do not control. The more useful question is what private AI looks like when a trader actually uses it. This is one workflow, described as it runs.

The workflow: document to platform data

The most common task Euclid AI handles is the one that consumes the most operator time: reading a trading document and getting its contents into the system. It follows standard manual capture logic and accelerates the execution. Three steps.

Ingest. You drop a trade document or trading email into the integrated chat assistant. Contracts, invoices, confirmations, across more than 50 formats.

Extract. The AI reads the unstructured file and populates the Deal Capture interface automatically. The fields a trader would otherwise type by hand appear filled.

Commit. You review the populated data and confirm it, committing the trade to the local CTRM domains.

The sequence looks like the manual process because it is the manual process, with the typing removed and a review step kept in place.

What each step protects

Two properties of that workflow matter as much as the time it saves.

The data stays inside the perimeter. Ingest and extract both happen on infrastructure the client controls, deployed on client-owned servers or inside Euclid's private Swiss data centre. The document is never transmitted to an external cloud for processing. For a file that might contain counterparty terms or pricing, that is the difference between using AI on it and not being able to.

A person stays in the loop. The AI populates. A trader commits. The Deal Capture review step is deliberate: the model does the mechanical work and a human confirms the entry before it reaches the book. Every step is traceable, which is what an audit trail requires.

The same pattern, across the platform

Document capture is the clearest example, not the only one. The same private architecture drives the other things Euclid AI does today.

Market data ingestion follows the identical logic. Rather than typing commodity prices from a daily PDF report, you let the AI parse high, low and mean prices and update the database. Same ingest-and-populate pattern, different document type.

Contextual support runs the other way. It answers questions instead of extracting data. New users ask the assistant how to do something in the platform. It responds with step-by-step guidance drawn from Euclid's technical documentation, so a new hire finds the answer without a call to support.

Three capabilities, one architecture. The model runs where the data lives.

The number, and why the architecture comes first

Euclid measured the average time to integrate a trading document by hand against the same task through Euclid AI. Manual entry: eight minutes for an operator to read a contract and enter it. Euclid AI: under ten seconds to return structured, populated fields. Both figures cover data entry only. Verification stays a separate human step and is excluded from both.

That 98% reduction is the visible result. The architecture underneath it is the reason the result is usable. A model that reads sensitive contracts is only an option if those contracts never leave the firm to be read. Run the extraction inside the perimeter and the speed becomes something a compliance-bound trading firm can actually deploy, rather than a demo it cannot risk on real documents.

That is the test worth applying to any private AI claim. Not how fast it runs, but where it runs while it does.

See how private AI works inside your CTRM.