Enterprise organizations are increasingly building AI agents. However, the results from these digital workers still vary widely. The difference often comes down to how enterprise knowledge is presented.
Jesús Barrasa addressed this challenge during a recent interview. Barrasa serves as the field chief technology officer of Gen AI at Neo4j Inc. He spoke with John Furrier of theCUBE Research during GraphSummit.
Barrasa explained that agents need more than just raw data access. They require access to enterprise meaning and institutional knowledge. Unfortunately, that context is not always captured or represented well.
As companies deploy more agents, they often encounter siloed development. Work completed for one application rarely carries over to the next. Barrasa compared this pattern to past reporting systems that generated conflicting information.
Teams frequently isolate a single problem and bake the required knowledge directly into individual agents through prompts and skills. When building a second agent, teams repeat the process. This mirrors reporting mistakes from seven years ago when different platforms delivered inconsistent results.
To fix this, organizations can implement a governed knowledge layer. This layer represents data assets, policies, concepts, and processes. It also helps humans audit how an agent reaches a conclusion. Explainability becomes critical as agents transition from simple conversation to active execution.
Barrasa noted that this context engine provides consistency and explainability. It traces data sources and highlights the exact elements used to generate an answer.
Knowledge graphs link raw data to the specific business concepts that agents require for interpretation. Companies do not need to build an exhaustive, enterprise-wide model right away. Instead, they can develop context across multiple applications step by step.
Barrasa advised starting with a single use case. When building the second use case, teams should align it with the first. Organizations construct the knowledge layer incrementally by identifying a use case, realizing value, and expanding. Large language models can significantly accelerate the construction of this underlying ontology.
Evaluating the return on investment for a shared knowledge layer requires new metrics. Companies should look beyond the performance of any single agent. They must measure how efficiency improves as subsequent agents are deployed.
Barrasa suggested tracking the cost efficiency of building agents two, three, and four because knowledge accumulates in the layer. Organizations must also monitor the cost of drift. This negative metric measures what happens when different agents produce diverging results or conflicting actions.
The complete video interview remains available through SiliconANGLE and theCUBE coverage of GraphSummit.
