Nvidia Says AI Agents Can't Tell a Tier From a Plan. Can Yours?
connecting agents to your data is not the same as giving them context
Shen Pandi
October 22, 2026
Most organizations assume an AI agent becomes useful the moment it is connected to their data. Plug it into the warehouse, the CRM and the shared drive, and it should be able to answer anything.
The larger problem is that access is not understanding. This week at GTC Berlin, Nvidia added an Agent Knowledge Foundation to its agent toolkit, built to give enterprise agents the company context they lack, The New Stack reported. Nvidia's VP of product for agentic AI, Adel El Hallak, said models can confuse business terms, treating a "tier" differently from a "plan," and struggle to find specifics in sources like a spreadsheet with many tabs.
That is the recurring work nobody budgets for: explaining to every new agent, assistant and analyst what "active customer," "revenue" or "churn" means at your company, and which table holds the version finance signs off on.
And it gets worse as you scale. Every new agent and every new connector adds another place where a definition gets guessed instead of known.
Why Can't an Agent With Full Access Answer a Simple Question?
Picture a sales operations agent asked how many enterprise customers renewed last quarter. It finds a "tier" field in billing, a "plan" field in the CRM and a "segment" column in a finance spreadsheet with six tabs. All three look like they mean enterprise.
It picks one. The answer comes back fast, neatly formatted and wrong by a few hundred accounts.
This is not an access problem. It is a context problem. The agent had every table it needed. It did not have the definitions, ownership and relationships that tell it which one counts.
The Cost Isn't Missing Data. It's Guessed Meaning
Nvidia's answer leans on knowledge graphs: an Auto-Ontology library that maps raw tables to business concepts like customers and orders, exposed to agents through an MCP server. In Nvidia's tests, its retriever models raised search relevance by 65%.
Better retrieval helps an agent find the right table faster. It cannot decide which definition your business agreed on, or who owns it. That knowledge has to exist somewhere first.
For most organizations, it does not. Gartner found that 63% of organizations either do not have or are unsure whether they have the right data management practices for AI, and predicts organizations will abandon 60% of AI projects unsupported by AI-ready data through 2026.
Before an agent can reason over your business, someone has to:
Agree one definition for each core term: customer, account, plan, revenue
Map every table and column that carries each term
Record which source wins when two of them disagree
Assign a named owner to every definition
Track which reports and agents depend on each one
None of these tasks are particularly difficult. The problem is that the answers live in people's heads, old wiki pages and chat threads, where no agent can read them.
How to Give Your AI Agents Business Context in 30 Days?
You do not need to redesign your data platform or launch a year-long ontology project to do this.
Within the first month, focus on the business terms your agents and dashboards use most: customers, accounts, products, plans and revenue.
Start by discovering where each term lives across your systems and resolving duplicate and conflicting records, so each entity has one trusted version. Then attach an owner and a plain-language definition to each one.
Next, expose this context through the same connectors and MCP servers your agents already use, so an agent checks the definition before it answers.
DataManagement.AI accelerates this process by automatically discovering metadata, mapping lineage across the modern data stack, identifying downstream dependencies, and creating a centralized knowledge layer that stays continuously updated as your environment evolves. It resolves duplicate and conflicting records, and ownership and policy travel with every entity, so when an agent needs to know what "enterprise customer" means, there is one governed answer waiting.
Instead of spending months documenting definitions by hand, your teams get a knowledge layer that agents read before they act.
Agent Accuracy Is Actually a Business Context Problem
The race to carry context to AI agents is moving fast. Nvidia is now working with Snowflake, Databricks and Palantir to do exactly that.
But every one of those layers depends on something only your organization can supply: agreed definitions, clear ownership and a known source of truth.
Organizations that scale AI agents successfully do not just give them more access. They give them the context to know what the data means.
Warm regards,
Shen and Team