Large language models (LLMs) understand the world, but they do not understand your world. They lack knowledge of your specific terminology, how your products relate to one another, which procedures apply to which configurations, or where your content boundaries lie. Without a structured foundation, AI retrieves by similarity, and in enterprise environments, similar is not the same as correct.
In a recent session, Seth Earley and Heather Eisenbraun of Earley Information Science explained how information architecture (IA) provides the semantic foundation that makes enterprise AI retrieval accurate, precise, and trustworthy. They outlined four core components of IA:
- Vocabulary: Normalizing terminology across systems to resolve inconsistencies.
- Taxonomy: Defining boundaries that keep retrieval paths narrow and precise.
- Ontology: Enabling AI to reason across connected concepts rather than returning isolated fragments.
- Metadata: Allowing AI to retrieve by meaning rather than similarity alone.
The session also introduced applicability logic, a mechanism that ensures AI returns not just a relevant answer, but the right answer—tailored to the correct product version, jurisdiction, role, and context. This, the speakers argued, is the difference between a consumer AI experience and a safe, enterprise-grade one.
Real-world examples from regulatory compliance and field service environments demonstrated how each component addresses a specific class of retrieval failure that prompting and model selection alone cannot fix. The takeaway: without robust IA, AI in the enterprise remains a risky gamble.