Enterprise Knowledge Graph Architecture
A permission-aware context layer that lets enterprise AI assistants answer questions about the organization itself, fast, securely, and at low cost.
Role: Senior Technical Program Manager
The problem
Enterprise AI assistants constantly need to answer questions about the organization they serve: who works on which project, which documents relate to which initiative, who is permitted to see what. Today, each AI application reconstructs that context from scratch on every question, issuing a series of calls across multiple systems, consuming tens of thousands of tokens per cross-system query, and taking several seconds to respond. Every application does this work independently and redundantly, and each one has to re-implement access control correctly or risk leaking information. It is slow, expensive, and difficult to keep secure.
My approach
I authored the strategic and product foundation for a shared organizational knowledge intelligence platform, and I am re-architecting it into a two-tier, metadata-first design. The inherited design stored an organization's content inside the knowledge graph itself; I am replacing that with a model that separates structure from content:
flowchart TD Q(["AI assistant query"]) --> G["Tier 1, Metadata knowledge graph
relationships only: who · what · access
Amazon Neptune + OpenSearch"] G -->|most questions answered here| L["Direct graph lookup
under 1 second · zero LLM cost"] G -->|content actually needed| F["Tier 2, On-demand fetch via MCP
retrieves only the identified item"] F --> P["Permissions enforced
at the graph layer"] L --> ANS(["Answer"]) P --> ANS
Two-tier, metadata-first retrieval: the graph answers most questions directly; content is fetched only when needed, after permission checks.
Metadata knowledge graph
A knowledge graph (Amazon Neptune + OpenSearch) that holds only the relationships among an organization's data, who, what, when, and who has access. Most organizational questions are answered by direct graph lookup in under a second, with zero language-model inference cost.
On-demand content fetch
When a question genuinely requires the contents of a specific item, the graph has already identified exactly which one, so the system retrieves only that, on demand, through emerging standards such as the Model Context Protocol (MCP).
Because the first tier stores structure rather than content, access permissions are enforced at the graph layer before any sensitive material is retrieved, security is a property of the architecture, not an afterthought.
Why it matters
This pattern addresses the organizational-context gap that the research firm Gartner projects will cause 40% of agentic-AI initiatives to fail by 2027. It applies to any large enterprise deploying AI assistants across fragmented data sources, not a problem unique to one company. The broader market has independently confirmed the stakes of this category: the enterprise-knowledge competitor Glean has reached a $7.2 billion valuation, and Foundation Capital has characterized organizational context graphs as a "trillion-dollar opportunity."