Neuro-symbolic AI combines neural models (perception, extraction, language) with symbolic systems (rules, logic, ontologies) so that each does the work the other cannot: neural components read the messy world, symbolic components guarantee consistent reasoning over it. In regulated domains it is the practical path to AI that is both capable and defensible.
Pure neural systems struggle to guarantee consistency; pure symbolic systems struggle to read reality. The hybrid keeps the guarantee where the liability lives — the determination — and applies learning where the labour lives: turning documents, drawings, and feeds into structured facts.
An ontology usually sits at the seam: a typed graph of the domain's entities and relationships that gives the symbolic layer its vocabulary and the neural layer its extraction targets.
This vocabulary is implemented, not aspirational — it describes how the MAIA substrate actually runs. See the platform or read the research.