In one sentence
You upload the documents — the co-pilot does the rest: it extracts the facts, links them into a relationship graph, tells the case as a living story, records every change with a diff and actively thinks along whenever contradictions or legal consequences arise. You no longer have to 'operate' the AI.
Why a fact graph beats RAG
Most legal AIs break documents into snippets and, for each question, retrieve the passages that look most similar (RAG). That is local and lossy: it finds what resembles the question — but not the connections within a case file. Yet those are exactly what decide matters: the witness in volume 1 is the defendant's father. Two statements contradict each other by two hours. One exhibit is missing, which is why the due diligence falls apart.
Andrej Karpathy put it pointedly: instead of dumping raw text or relying on fuzzy search, you build a structured, evidenced knowledge artefact once — like a wiki: entities with a stable identity, explicit relationships, a chronology, flagged contradictions — and you give the AI that as context. That is precisely what we do for the case file.
The benefit is concrete:
- Global coherence — the AI sees the whole case as a structure, not k random snippets.
- Relationships as the substrate for reasoning — edges such as 'contradicts', 'witness for' or 'evidence against' are the primitives of legal reasoning; the AI can traverse them.
- Stable identity — 'Mr Müller' across five documents is ONE node, not five unconnected hits.
- Compression with a source — the graph plus a curated summary fit into the context, and every fact carries a pointer to its source (document, page, quotation).
Facts, nodes, edges — and you stay in control
From every document the co-pilot pulls structured facts: parties, organisations, dates, deadlines, amounts, clauses, evidence. These become a relationship graph with more than 25 legal edge types — displayed graphically with directional arrows, colour-coded by meaning (supporting, contradicting, correcting) and a click through to the source in the document.
Importantly: every relationship the AI proposes is at first merely a lead to review (dashed). You confirm or reject it with a click — only then does it count as established. That preserves both evidential traceability and liability certainty.
The living case narrative — and its history
Instead of a static summary, the co-pilot tells the case document by document and as a chronology: 'the claim is about …, the defence is about …, the reply is about …'. The depth is adjustable per case — from a single page to a detailed, multi-page version.
When a new filing or exhibit is added, the current picture updates automatically — and you see the diff: what has changed? Which value was corrected (e.g. claim €12,000 → €12,345)? Which new relationships, which new or closed gaps? Using a time slider you can even call up any earlier state again: 'this is how the case stood when the claim was filed.'
The AI that thinks along
The co-pilot reviews the case like an experienced colleague:
- Timing and logic contradictions: 'the witness states they were present at 14:00, but according to the record only arrived at 16:00 — so they cannot have observed the events.'
- Legal implications under the right law: for a German case, say, along the lines of the German Code of Civil Procedure (ZPO); for a French one, along the GDPR — the AI applies the case's jurisdiction, never German law across the board.
- Missing documents: 'Exhibit K5 is referenced in the statement of claim but is not on file — please submit it and I'll then complete the case analysis.'
Knowledge per client — and automatic conflict-of-interest alerts
Facts are accumulated per case and per client. If the same opposing party turns up in another client's case, the co-pilot warns of a possible conflict of interest — without disclosing any of the other case's contents. Knowledge stays separated by client; professional confidentiality is preserved.
Everything feeds into chat and pleadings
The co-pilot feeds the graph, the narrative, the history and the inferences it has drawn straight into the AI assistant and into drafting pleadings. Before drafting, it shows its basis — which facts, which sequence of events, which inferences — and asks for confirmation. If something is missing, it flags it as a placeholder to query with the client, rather than bothering you with routine questions. The result is a finished pleading, in the language of the case and in the style of the relevant jurisdiction.
Built European and multilingual
Language and law are kept separate: the case language determines the output (German, English, French, Italian, Spanish, Dutch, Polish), while the jurisdiction determines the law to apply. Both come from the engagement or the firm settings — and can be overridden per case.
Full auditability
Every step lands in the log, every write action in the audit ledger — with undo. Relationship reviews, corrections and state versions are documented without gaps. The co-pilot works in the background, EU-resident and with PII protection — but the decision always stays with you.
Conclusion
An AI that knows how the case has evolved, understands all the facts and their connections, spots contradictions and legal consequences on its own and drafts pleadings from them — that is the difference between 'analyses a PDF' and 'thinks with the whole case file'. That is exactly what the fact graph is built for.
