AI and notarial practice: how the Caen Notarial Chamber transforms its work with MarIAnne

In brief. The Caen Notarial Chamber (CNCAC) deployed MarIAnne, a domain-specific document AI under expert supervision. Scope: 415 notaries, about 9,000 knowledge segments, and more than 12 months in production.

---

Context

The Caen Court of Appeal Notarial Chamber (CNCAC) oversees a regulated legal profession facing major change. Since the 2015 reform, headcount has nearly doubled: 255 out of 415 notaries were appointed after 2015, and 70 offices were created out of 194.

That shift significantly increased support demand: more deontological, regulatory, and operational questions, with rising response times for the expert team.

Three observations triggered the project:

• Doctrine was fragmented across rules, circulars, internal sheets, and tacit expert knowledge.
• Request volume could no longer be absorbed without delays.
• Knowledge transfer remained fragile in the face of retirements and turnover.

The goal was not to "launch a chatbot," but to turn critical documentary assets into an on-demand organizational memory.

---

The problem to solve

A generic AI assistant was not reliable enough in this setting:

• risk of hallucinated sources or outdated references;
• difficulty prioritizing local rules when they supersede national ones;
• terminology ambiguity in a strict normative context;
• no structured learning from expert corrections.

The real need was to answer quickly, with the right sources, while distinguishing local vs national and valid vs repealed texts, and continuously improving with humans in the loop.

---

Our solution

We designed and deployed a full domain document assistant, owned by CNCAC, built around four core components.

1. A domain knowledge graph

The Chamber's corpus (national professional rules, CSN circulars, collective agreement, legislation, internal guides and sheets) is ingested, segmented, enriched, and indexed in a knowledge graph on Neo4j.

Outcome: nearly 9,000 structured, dated, and sourced segments, each traceable to its original legal or procedural text.

2. A precision-oriented hybrid RAG

Instead of naive vector-only retrieval, MarIAnne combines graph retrieval with LLM synthesis. The engine applies re-ranking based on source hierarchy and text status, then forwards only qualified passages to the model. Responses are systematically sourced.

3. The Tribunal: continuous human control

This is the centerpiece. A Chamber expert reviews generated answers, classifies errors (wrong source, obsolete text, ignored local rule, ambiguous wording, etc.), and applies corrections with veto power.

This is not passive annotation. It is operational co-governance.

4. A semantic improvement loop

Each correction becomes a structured, traceable knowledge object (author, date, act type, graph targets, rationale). The knowledge base evolves in a versioned, reviewable, and continuous way alongside doctrinal updates.

---

What changes for the Chamber

• Response time. Questions that previously needed calls and email exchanges now get a first answer in seconds.
• Consistency. Two notaries asking the same question receive coherent answers grounded in the same sources.
• Knowledge preservation. Expert know-how is formalized, versioned, and traceable, so knowledge remains within the organization.
• Faster ramp-up. New entrants gain immediate access to accumulated doctrine.
• Quality steering. Source relevance becomes a trackable metric rather than a black box.

---

Differentiating principles

This case exemplifies GenieFactory's agentic transformation approach for high-reliability environments:

• Humans remain in control. AI proposes, experts decide.
• The client owns the system. Data, graph, sheets, and corrections belong to CNCAC.
• Personalization is native. The assistant reflects local rules, domain vocabulary, and regional constraints.
• Knowledge stays alive. It is tracked, versioned, and continuously enriched.

---

Capabilities delivered

This project required end-to-end capabilities:

• knowledge engineering (modeling, Neo4j graphs, ingestion pipelines);
• advanced RAG (hybrid architecture, domain re-ranking, hierarchical retrieval);
• product design for a critical use case (feedback-focused conversational UX);
• infrastructure and operations (monitoring, orchestration, continuous updates);
• domain enablement (expert workshops, quality governance).

---

A replicable model

MarIAnne now runs at interdepartmental Chamber scale in Caen. The setup is transferable to other regulated professions with similar challenges: knowledge transfer, scaling demand, and ensuring consistency of practice.

Is your domain knowledge a durable asset, or a hidden loss risk?

Discuss your project →