NVIDIA Adds a Business Data Layer to Enterprise AI: Moving from Information Retrieval to Prediction and Decision-Making

AI assistants in enterprises are already capable of finding answers in documents, but when faced with questions such as “Which orders might be delayed?” or “How should this revenue be calculated?”, simply retrieving information and expressing it fluently is not enough. They need to understand the relationships between data points, the business implications of metrics, and what actions they can take.

October 21, 2026,NVIDIA Announcement of Expansion Agent Toolkit...By integrating the Agent Knowledge Foundation, it combines retrieval models, structured data prediction models, business semantics tools, and accelerated computing libraries. This provides a more comprehensive technical path for enterprise AI: first, identify the correct information and clarify its meaning; then, perform predictions and optimizations.Official Announcement

Just because data can be retrieved doesn't mean the business implications are the same.

Suppose a procurement specialist asks, “Which suppliers are most likely to affect deliveries next month?” The system may need to cross-reference orders, inventory, logistics records, and supplier information simultaneously. Even if the query can be executed, the conclusion will vary depending on whether “delivery time” refers to the promised date or the actual arrival date, and whether canceled orders are included in the statistics.

These problems are difficult to solve solely through expansion...Large model parametersSolution. First, a company needs a shared set of business definitions that allow customers, products, orders, and metrics across different systems to be mapped to one another. Technically, these objects and relationships can be expressed using an “ontology”—a relational map that enables machines to understand business concepts.

NVIDIA’s Auto Ontology is responsible for establishing this layer. According to its official project description, it organizes business semantics from enterprise metadata, harmonizes definitions across different systems, and, based on certified definitions and data lineage checks, processes queries and provides the sources of results. It complements existing semantic models rather than requiring enterprises to start from scratch.Official Auto Ontology Project

The key lies in who performs the “validation.” NVIDIA explicitly states that the automatically generated ontology serves as a starting point, which must then be tested, revised, and validated by business experts. Finance, procurement, or operations personnel remain responsible for ensuring the consistency of metrics; while the tool reduces the workload associated with organizing and integrating data, it cannot decide on their behalf what constitutes the correct business definition.Agent Knowledge Foundation

From Searching for Information to Planning the Next Step

This foundation layer is not a new model that handles everything. Nemotron Retriever focuses on document processing and retrieval; Auto Ontology provides business semantics; and Kumo is geared toward predictions on tables, relational tables, and time series. The underlying CUDA-X libraries handle data processing, relational analysis, machine learning, vector search, and optimization calculations. Enterprises can select components as needed, without having to adopt a new, comprehensive platform.Components and Workflows

The significance of Kumo is particularly easy to overlook. According to the developer documentation, Tabular is designed for single-table forecasting, Relational leverages relationships between multiple tables, Forecast is geared toward time series such as demand, and Anomaly is used to detect anomalies. This gives AI assistants the opportunity to connect “what happened in the past” with “what might happen next,” rather than simply searching for descriptions in pre-existing reports.Kumo Official Introduction

Taking delivery as an example again, a well-designed system would first retrieve the contract terms, then read the relevant orders, generate a delay risk estimate, and finally propose adjustment plans subject to production capacity and shipping constraints. This is a scenario simulation based on component capabilities and does not represent the complete delivery system that NVIDIA has publicly validated. Risk predictions are not facts and certainly cannot directly replace procurement decisions.

Why Integrate with an Existing Data Platform?

NVIDIA’s announced implementation approach reflects this: its supply chain organization leverages these capabilities through Palantir’s Foundry, Ontology, and AIP; companies such as Databricks and Snowflake are also integrating relevant components into their respective environments. What’s more noteworthy here is how these capabilities are integrated into existing enterprise systems, rather than the length of the partner list.Enterprise Adoption

For companies that have already invested in data warehouses, permission systems, and analytics tools, building a separate AI data silo would increase the burden of synchronization and maintenance. Integrating semantic, search, and predictive capabilities into existing workflows may make it easier to leverage existing investments. However, integrating these components still incurs costs related to computing power, inference, data preparation, and operations and maintenance; open source does not mean free deployment.

What Issues Should Be Tested Before the Official Launch?

First, check whether the business scope is consistent. If you rephrase the same question, does it still yield the same statistical scope? Can the answers be traced back to specific data, versions, and calculation methods? For forecasting tasks, you should also validate the results using data from later time periods to avoid introducing known outcomes into the test.

Second, distinguish between read and action permissions. The ability to analyze supplier risk does not mean that orders should be automatically modified. OpenShell in the Agent Toolkit provides runtime access and behavioral boundaries, but companies must still define which actions are allowed to be performed automatically and which require approval from business managers.Runtime and Governance Notes

Configurations also need to be checked separately. The official Auto Ontology repository notes that when a data source connection is added via the interface, connection objects—including passwords—are stored in plain text in the database if Vault is not configured. This is a risk specific to this deployment method and should not be obscured by general descriptions such as “enterprise-grade” or “secure runtime.”Connection Management Instructions

Finally, the total cost of each valid result should be evaluated: how much time is saved on manual verification, how many erroneous recommendations are generated, and how much rework is required. The performance improvements listed by the official sources cannot be directly converted into a company’s operational returns.

This also determines what kind of teams it is best suited for: companies that already have a data platform, business owners, and engineering maintenance capabilities—and that genuinely need cross-table analysis or forecasting—will be able to evaluate each component more specifically. Small teams that simply want to answer questions about documentation, however, do not need to deploy all the tools at once just to follow a “complete architecture.” It makes more sense to first identify whether the issue stems from retrieval, data consistency, or forecasting, and then decide which layer to address, rather than simply stacking components on top of one another.

One aspect of NVIDIA’s latest expansion worth noting is its extension of enterprise AI reliability to the data and business definitions themselves. While model capabilities remain important, as AI assistants begin to influence orders, inventory, and resource allocation, the ability to explain “how this conclusion was reached, who approved it, and what actions are permitted” will become a more practical competitive advantage.