How a scalable intelligence layer turns enterprise data into production A
SiliconANGLE Jonathan Anthony ● Covered by 3 sources
Microsoft's Jeevan Pathuri says knowledge graphs are the missing layer turning AI pilots into real production systems. Connected data, not raw model power, is what makes agents actually reliable.
Based on reporting by SiliconANGLE, Jonathan Anthony — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
There's a quiet realization spreading through enterprise AI teams: the model isn't the hard part anymore. Getting a large language model to sound smart is table stakes. Getting it to know your business — your suppliers, your product hierarchies, how one system's data relates to another's — is the actual work, and it's the part most companies have been skipping.
Jeevan Pathuri, vice president of software engineering at Microsoft, laid this out plainly in a conversation with theCUBE's John Furrier. His argument is that the real intelligence in enterprise AI doesn't come from the model itself but from the relationships buried in a company's own data. Model that data as a graph — nodes, edges, connections — and suddenly a language model can answer nuanced questions instead of just running a glorified SQL lookup.
Microsoft tested this on its own supply chain. By representing a bill of materials as a graph rather than a pile of disconnected tables, Pathuri's team built somewhere between 10 and 15 agents in production within a few weeks. Compare that to the old way — building each agent from scratch, rediscovering the same metadata and relationships every single time — and the difference is stark. An agent that once took three or four weeks to develop, he said, can now be built in a matter of days once that semantic layer exists.
That's a big claim, but it lines up with where money is flowing. The enterprise knowledge graph market is projected to grow from roughly $1.9 billion today to nearly $10 billion by 2032, and startups like Jedify are raising capital specifically to give AI agents the kind of structured business context that generic models lack on their own. TheCUBE Research's own 2026 predictions echo this, naming knowledge graphs and semantic layers as central to whether enterprise AI deployments actually work.
But Pathuri isn't selling this as a shortcut. He's blunt that there's still hard work involved — building a clean, connected data layer doesn't happen by magic, and humans still need to stay in the loop with evaluations and production tracing. The pitch isn't that graphs replace effort. It's that they let companies spend that effort once, on infrastructure, instead of over and over on every new agent they try to ship.
My take — AI-written commentary, not fact-checked reporting
Every AI cycle eventually produces a moment where someone points out that the fancy model was never the bottleneck — the boring data plumbing was. Knowledge graphs sound unglamorous next to talk of frontier models, but Pathuri's numbers, going from weeks to days per agent, are the kind of practical payoff that actually survives contact with a budget committee. Expect more Microsoft engineers, and more well-funded startups, to quietly build businesses on making enterprise data legible to machines rather than making machines smarter in the abstract.
Read more about this at: SiliconANGLE