Ask an AI agent a question it cannot answer, and it rarely admits it. Instead, it guesses, confidently, in complete sentences, and the answer often sounds right even when it is not. Enterprises call this “AI hallucination”. But the real failure usually happened earlier, when the agent went looking for an answer and could not find the right data, or found three versions of it that did not agree.

That gap gets more expensive as AI moves from a single chatbot to a team of agents working a task together. A chatbot answers one question and stops. An agent reads a ticket, retrieves a document, queries a database, calls a tool, maybe hands off to a second agent, and only then responds. Every one of those steps depends on data the agent can actually find, trust and act on. One weak link anywhere in that chain, and the error travels downstream with it. Now we are all hallucinating.

Agent Infrastructure Is Not Just Compute

Most conversations about scaling enterprise AI start with GPUs: more of them, faster ones, bigger models behind them. All useful, but none of it answers the harder question: how does an agent know which data is current, who is allowed to see it, what a given field actually means, and where an answer came from if someone needs to check it later? That is the data plane, and it is turning into its own category of infrastructure, as foundational to agentic AI as a network is to the internet.

The stakes rise after more than one agent touches the same information. A support agent, a finance agent and an inventory agent might all pull from the same customer record, but if one system calls it a “client” and another calls it an “account,” two confident agents can reach two different conclusions from the same underlying fact. Scale that across dozens of agents built on different tools by different teams, and disagreement becomes the default, not the exception.

This is the specific problem Dell Technologies is addressing with the newest additions to the Dell AI Data Platform with NVIDIA, the data context layer of the Dell AI Factory with NVIDIA. A unified semantic layer gives structured and unstructured data the same business meaning wherever it shows up, through shared definitions and a searchable glossary, so “client” and “account” resolve to the same thing instead of two. An enterprise knowledge graph then maps how that data actually connects, using metadata, lineage and query history to keep tuning itself as the business changes, so an agent chasing one odd sensor reading can trace it to the machine, its repair history, the supplier batch and the orders now at risk. Knowledge agents sit on top of that graph, each one grounded in a defined slice of context, with the customer setting the rules for what it can see, what quality bar it has to clear, and what it is allowed to spend. None of it has to leave the building. The context stays inside the platform the customer already runs.

Structured And Unstructured, Together

Most of the attention in enterprise AI, understandably, goes to unstructured data. It makes up somewhere between 80 and 90 percent of what organizations generate, and most of it has been effectively invisible to AI until recently. Dell’s Data Orchestration engine already does the heavy lifting there, documents, images, video, PDFs, pulling that content into AI-ready pipelines. But a transaction table and the PDF explaining that same transaction are not two competing sources of truth. They are two halves of one answer. The newer move is pulling structured database tables into that same pipeline, so an application can combine the fact sitting in the ledger with the detail buried in the document, rather than being forced to pick one. Alternatively, treat structured data as an afterthought, and an agent will answer with complete confidence and half the picture.

Four Agents, One Foundation

Here is what that looks like on a factory floor with Dell AI Data Platform with NVIDIA. A technician snaps a photo of an unfamiliar error code, and four agents start working at once: one searches manuals and service bulletins for the fault, one estimates the part’s remaining life from sensor history, one pulls the exact repair clip from a video archive, and one checks warehouse stock for the replacement part. Each agent draws on a different engine: search, processing, analytics and orchestration. All four draw on the same governed foundation. The technician gets one coherent answer instead of four tools to reconcile by hand, and the plant avoids the downtime that used to depend on one retiring expert’s memory.

None of this holds up if the data path itself is slow. A knowledge graph that takes minutes to traverse, or a search index that cannot keep pace with a GPU cluster sitting idle and waiting, turns a sound architecture back into a bottleneck. That is why acceleration has to run the length of the stack, from storage through the data engines to the model itself, which is where Dell’s engineering work with NVIDIA picks up the story.

Enterprises are not short on data. They are short on data their agents can trust fast enough to act on. That is an infrastructure problem before it is a model problem, and the organizations that solve it first will have agents that get work done instead of just sounding like they did.