Technology
How a figure gets from the paper to the document.
Specialised agents coordinated by a single orchestrator, a knowledge graph for every engagement, a second model that looks for the error before delivery.
In five steps
The path of a figure, from the client’s paper to the document.
1 of 5. The papers
It starts from the client’s papers.
Financial statements, contracts, Central Credit Register, term sheets: even hundreds per engagement, in different formats. A figure taken from these papers carries with it the passage it comes from.
Context, control, trail
An agent works within a perimeter.
It has the engagement’s papers and your method in front of it, it can only do its own task, and it leaves a trail of what it does. The models change; context, controls and trail stay.
Context The engagement’s papers and your method.
The agents start from the client’s papers, linked in the knowledge graph, from your example document and from your analyst’s corrections; for a bank, also from policy, RAF and approval powers. Where data is missing, the section is flagged instead of being filled in.
The engagement’s knowledge graphPeople, companies, contracts and guarantees, linked to the papers they come from.
The agents consult it so as not to miss links and inconsistencies, even in the largest files. An example from a financing engagement:
| Entity | Relation | Entity | Source |
|---|---|---|---|
| Director | represents | Company | Minutes, p. 2 |
| Company | signs | Loan | Contract, art. 3 |
| Mortgage | secures | Loan | Registration note, p. 4 |
| Mortgage | encumbers | Property | Land registry search |
Learning from correctionsYour flags become instructions for later engagements.
A flag left at the point of the document it refers to becomes an instruction, an example and a check for the agents on later engagements. Your content trains no model.
Control One task, the tools for that task.
The engagement’s process sets the order, and whatever fails the checks goes back. The agents see only your organisation’s papers and do not write into your systems. You decide what goes into the document.
Orchestrating the agentsThe work split into steps that can be checked, each with its own record.
An agent performs one precise task with the right tools. Orchestration coordinates them from the start to the end of the engagement.
The right model for each taskDifferent tasks, different models. No single provider.
Reading a scanned contract, rebuilding a set of accounts, writing a report: each task goes to the model that does it best. We test new models and adopt them without changing the way you work.
Adversarial reviewWhoever wrote a text struggles to see its flaws.
Every draft goes to a second AI model, different from the one that wrote it, with a single task: find what does not hold. Each finding is checked against its source, then corrected or rejected with a reason. Before delivery, the consistency between sections, the reconciliations and the completeness are checked.
Trail What happened, and who keeps it.
Every step stays recorded, and what the engagements produce stays with your organisation.
The trail of every stepWhich agent, on which paper, with what outcome.
Every step of the agents and the people stays recorded, and the full trail arrives with the document. The knowledge graph, the wiki and your standard grow from your engagements and stay with your organisation: at the end of the contract you export them.
The data stays in EuropeDocuments in Milan, AI processing in the European Union, no retention.
The rest is on the Security page, with the status of each measure.
Glossary
A small glossary.
Tap a word: its meaning appears alongside.
In ArgusOne each uses a for one precise task. The coordinates them and keeps a record of every step. The links people, companies and contracts. Before delivery the hunts for . And with the model provider retains nothing.
A program that uses an AI model for one precise task, with the right tools: reading a paper, extracting a figure, writing a section.