Frédéric Valentin designed the AI tax MCP used by the members of the Cercle des Fiscalistes, an independent French think tank of tax practitioners who advise on cross-border wealth. The starting point is familiar: language models cite repealed articles, invent thresholds and mix up jurisdictions. Pre-trained weights are out of date the day they ship, and tax law moves with every Finance Act.
The system rests on a Neo4j knowledge graph that indexes the tax law of nearly twenty jurisdictions — statutes, treaties, administrative doctrine, case law — and on an MCP server that any assistant can connect to. The graph does not store the law: it says where the law is and which version applies. The official text is read at answer time, and the URL is the citation.
The session covers how the law is modelled in the graph, how the MCP tools are organised for an agent, and how full-text search and graph traversal work together to find a rule and follow its dependencies. It closes with the lessons of going from one jurisdiction to nearly twenty, and a short demonstration: one question, one cited answer.
Speaker
Frédéric Valentin
Chair of the AI committee, Cercle des Fiscalistes
Frédéric Valentin is a wealth-planning practitioner in SGPB. A member of the Cercle des Fiscalistes, he chairs its AI committee and built the Neo4j knowledge graph and MCP server that let its members' assistants cite tax law instead of guessing it — nearly twenty jurisdictions, in production since July 2026. He teaches international tax law at HEC Paris and Sciences Po, and is a member of the IACF (Institut des Avocats Conseils Fiscaux) and of the Estate Planning Skill Group. Le Revenu named him one of France's 12 leading tax lawyers (Prix du Revenu 2026), and Neo4j a Featured Community Member in July 2026. He builds his own tools, in Python and Swift.
