couche sémantique

Jamin Ball's ARR example is the sharpest statement of the problem: inside a single company, sales, finance, accounting and legal can each compute annual recurring revenue from different definitions and conflicting data sources. Humans arbitrate that ambiguity as they work. Agents do not. In Ball's December 2025 note, a quote-to-cash workflow that picks up the wrong price list, the wrong contractual term or a stale ARR figure will then automate the wrong thing, confidently, at every step downstream.

The couche sémantique is what both sources reach for: a shared understanding of the enterprise (entities, relations, definitions, rules) that encodes what data means and which source takes precedence. Ball puts it in the data warehouse and the lakehouse, which he recasts as "truth registries" supplying the semantic layers and governance frameworks agents need to operate safely. Padhi and co-authors at McKinsey place it one level up, inside the AI operating system: a layer of ontologies and the Knowledge graph, resting on existing infrastructure and data, alongside deterministic analytical models called by agents, agent orchestration and an experience layer.

The gap both are writing against shows up in McKinsey's State of AI 2026 survey: 40% of large companies have deployed agents at scale, 6% credit AI with at least 5% of EBIT. Fixing semantic governance sits in McKinsey's opening recommendations, next to funding the shared AI OS and redefining decision rights. Ball's own conclusion is that a system of record is a canonical source of truth, which changes what a company actually has to build.