Layered data architecture turns enterprise data into a system of intelligence

The knowledge graph is becoming a crucial foundation for enterprise artificial intelligence as companies strive to convert disorganized data into reliable answers for decision-makers. As cloud-native stacks now transition to AI-native, a new layered data architecture is emerging - one specifically designed to provide models with the context they require to reason, not just retrieve data.

This shift is occurring across a diverse data landscape including lakehouses, operational databases, and customer profile stores, where the key challenge lies in precisely describing an enterprise's information so AI can understand it. Graph technologies are increasingly being used as the connective tissue for AI agents and GraphRAG architectures, according to Tristan Baker, senior director and head of data architecture at Salesforce Inc. "It's becoming that critical piece that helps tie the end agentic experience you want to deliver," Baker explained.

"It's like the glue that stitches what the customer or the person is asking to the data in the context needed to answer that question." At the Neo4j GraphTalk event, Baker discussed with John Furrier how knowledge graphs, ontologies, and layered data architecture are transforming the quest for an enterprise system of intelligence.

The ultimate goal for leaders is a conversational interface that delivers trustworthy answers in seconds. However, achieving this speed requires more than a single database, as different questions necessitate different retrieval structures. Baker emphasizes that delivering answers involves more than just providing the number; it requires explaining the reasoning process and showing all the connections to relevant information.

He frames the solution as a stack, with foundational layers such as lakehouses, operational, and time-series databases, each optimized for specific query types. Above these layers, a metadata layer tracks where the truth about a customer resides across multiple copies and maps business terminology to underlying columns. The graph manages the relationships and contextual information that link these elements together.

Despite the advancements, Baker acknowledges that solving this problem is far from complete. He notes that like any system, even with the best technology, "garbage in, garbage out" applies. Without careful attention to the content being fed into the system, useful results cannot be expected. One of the most under-addressed challenges, Baker adds, is governance.

As metadata moves up to a higher semantic layer, traditional access control mechanisms residing within individual databases are no longer sufficient. Baker points out the need for master data management and governance to work together. He explains that, beyond having a semantic description of the data, there also needs to be a semantic description of access policies.

Legal teams, for example, will not simply provide guidelines on data protection within Postgres. Instead, they will specify who should not be able to access certain data.