Building Reliable AI Agents With Knowledge Gardens and MongoDB

What happens when an enterprise AI agent can retrieve thousands of data points but cannot understand the customer, decision, or business moment in front of it?

In this episode of Tech Talks Daily, I welcome back Boris Bialek, Vice President of Industries and Global Field CTO at MongoDB. We examine why the enterprise AI conversation has become more professional as organizations move beyond demonstrations and begin putting agentic systems into production.

Boris argues that many companies do not have a shortage of data. Their problem is turning scattered data into information and then into usable knowledge. A bank balance is data. A complete view of a customer’s relationship with the bank is information. Recognizing that the customer is currently researching a mortgage and may need assistance within the next 20 seconds is knowledge.

This distinction leads to Boris’s concept of a knowledge garden. Structured records, unstructured content, live signals, conversations, and business context are organized around a customer or outcome. Different departments can access the parts relevant to their work while AI agents receive the context needed to respond quickly.

We also discuss integration debt. Boris recalls one system that required 18 seconds to assemble a customer view and says many enterprises are working with approximately 40 primary data sources. An agent can spend so much time coordinating access across APIs, caches, and applications that the business problem becomes secondary.

Trust becomes equally important once an AI agent can act. Boris introduces two measures: the agent confidence score and the business risk score. The first evaluates whether an agent’s output appears reliable based on its data, behavior, and context. The second considers the consequences of allowing that decision to proceed automatically.

Together, these scores can help organizations decide which actions should pass automatically, which need further machine validation, and which should reach a human reviewer. Boris also explains why data lineage and complete audit trails must be designed into production systems from the beginning.

For teams beginning this work, his advice is practical. Choose one business outcome, connect two or three relevant data sources, create a working prototype, and involve business and technical leaders in the same conversation. The goal is to demonstrate how data, context, confidence, risk, and human review work together before expanding the system.

Does your organization have an AI data problem, or does it have a knowledge and context problem? Listen to the conversation and share your thoughts with me.