Neo4j Makes Case for Shared Knowledge Graphs as AI Agent Context

Barrasa said organizations have become better at building agents, but their results still vary widely. The difference often lies in how enterprise knowledge is presented. “You have to give agents not only access to your data, but also to your meaning, to your enterprise knowledge,” he said. “That’s unfortunately not always well captured … [or] well represented.”

He described a familiar pattern: teams identify a single problem and build the knowledge an agent needs inside the agent through prompts and skills. When they build a second agent, they repeat the work. “So, we’re repeating the errors that we made seven years ago when we were building reports in different platforms and getting inconsistent results,” Barrasa said.

A knowledge layer, as Barrasa described it, is a governed representation of an organization’s data assets, concepts, policies and processes. It can serve as a context engine for agents, providing consistency and explainability. “This knowledge layer, this idea of capturing your enterprise knowledge and using it as the context engine for your agents, not only gives you the consistency that we were talking about before, it gives you the explainability,” he said. “That’s the source of my data. These are the elements that I use to produce the answer.”

Knowledge graphs connect data to business concepts and relationships that agents need to interpret it. Barrasa said organizations do not need to build an enterprise-wide model at the outset. “You want to start with one use case. And when you build use case two, you have to try to align it to number one,” he said. “That’s how you build the knowledge layer. You build it incrementally. It’s ‘identify use case, realize value and then build from there.’ The construction of the ontology is something that [large language models] can accelerate significantly.”

He also suggested a broader way to measure return on investment than the performance of any single agent. Organizations should assess how the construction of Agent 2, Agent 3 and Agent 4 becomes easier as knowledge accumulates in the layer. “Another [metric] is … a negative metric: What’s the cost of drift? What happens when two agents return diverging results or act in different ways? What’s the cost of reconciling these results?” he said.

The interview was conducted by theCUBE Research’s John Furrier and was part of SiliconANGLE’s and theCUBE’s coverage of GraphSummit. theCUBE is a paid media partner for GraphSummit. Neo4j, the sponsor of theCUBE’s event coverage, and other sponsors do not have editorial control over content on theCUBE or SiliconANGLE, according to the disclosure accompanying the interview.

Editor's Summary

Neo4j executive Jesús Barrasa argued at GraphSummit that enterprises need a shared, governed knowledge layer to give AI agents consistent context and explainable answers. He said knowledge graphs can be built incrementally across use cases, with LLMs accelerating ontology construction. Barrasa also urged organizations to measure avoided drift and reduced effort as new agents are added.