AI CX AutomationDigital transformationKnowledge management
From Knowledge to Trusted Knowledge: The Governed Knowledge Pipeline for the Agentic Enterprise
Executive summary
NVIDIA has now named knowledge as one of the five elements every business agent needs. Knowledge that is retrieved is not yet knowledge that can be trusted, and that gap is where agents fail inside regulated enterprises.
At GTC Berlin on October 21, 2026, Jensen Huang defined five elements that comprise an AI Agent for business: a model, a harness, a sandbox, knowledge and tools (NVIDIA). He launched the NVIDIA Agent Knowledge Foundation to connect agents to enterprise data and said that company knowledge is what turns agents into super agents.
We propose a reference architecture that completes the knowledge element. NVIDIA’s Agent Knowledge Foundation finds, relates and predicts. The eGain AI Knowledge Hub sits directly downstream: it takes that output, cleans and collates it, finds the gaps, puts it under human supervision and governance, assigns accountability, and only then feeds it to agents. We call the result the Trusted Knowledge Pipeline.
The outcome for an enterprise is simple to state. Agents built on NVIDIA infrastructure act on knowledge that is verified, owned, current and auditable, so the business can stand behind what its agents say and do.
The Moment: Knowledge Becomes One of the Five Elements of an Agent
For the first time, the company that supplies the infrastructure under nearly all of AI has said that agents cannot be effective in a business without that business’s knowledge.
In his GTC Berlin keynote, Huang described an agent as more than a model answering questions: it draws on knowledge and acts through tools (NVIDIA keynote blog). The five elements he named are:
| Element | Role in the Agent |
|---|---|
| Model | Reasons and generates |
| Harness | Orchestrates the agent’s work and memory |
| Sandbox | Isolates the harness (NVIDIA OpenShell) |
| Knowledge | Gives the agent context from the company’s structured and unstructured data |
| Tools | Lets the agent act, including CUDA-X simulators, solvers and data engines |
Three points from the Huang keynote matter for our proposal:
- Knowledge is a pillar, not a plug-in. NVIDIA introduced the Agent Knowledge Foundation as a set of models and libraries for connecting agents to enterprise data (NVIDIA).
- Proprietary knowledge is the differentiator. Huang said that giving agents access to company knowledge is what turns them into super agents (BW Businessworld).
- Control is a first-class requirement. Huang argued that agents should run with minimal privileges under constant monitoring, and that capability and safety must advance together (Digit).
NVIDIA has secured the sandbox. We argue that the same discipline must apply to the knowledge element: an agent with perfect isolation that acts on wrong, stale or unowned knowledge is still a liability.
Why Retrieved Knowledge Is Not Yet Trusted Knowledge
Retrieval answers “what did we find?” A business needs an answer to a harder question: “is this what we stand behind?”
Even coverage of NVIDIA’s launch notes the gap: an agent can retrieve technically accurate information and still misread its significance, and it can produce unreliable recommendations when the underlying data is incomplete or outdated (IntelligentHQ). In customer-facing and regulated work, five failure modes recur:
| Failure Mode | What It Looks Like | Consequence for the Agent |
|---|---|---|
| Conflict | Two policies, a product sheet and a call-center script give different answers | The agent picks one, inconsistently |
| Staleness | A superseded rate, fee or clinical guideline still ranks highly | Confident, out-of-date answers |
| Gaps | Customers ask questions no approved content answers | The agent improvises |
| Unverified content | Drafts, emails and wiki pages sit beside approved policy | Unapproved language reaches customers |
| No owner | Nobody is accountable for a given answer | No audit trail, no remediation, no defensible position |
None of these is a retrieval-quality problem; better embeddings and reranking will surface a wrong document more precisely. They are knowledge-management problems. They are solved by curation, reconciliation, human judgment and accountability, which is the discipline eGain has built for more than two decades.
NVIDIA’s Agent Knowledge Foundation: Search, Analyze, Predict
The Agent Knowledge Foundation turns enterprise data into material an agent can use, without moving the data or replacing existing workflows (IntelligentHQ). Its components group into three functions:
| Function | NVIDIA Components | Output |
|---|---|---|
| Search | Nemotron Retriever (embedding, retrieval, reranking); cuVS vector search | Ranked passages and documents |
| Analyze | Auto-Ontology (raw data to business concepts and relationships); cuGraph, cuDF | Entities, relationships, ontology |
| Predict | Kumo Tabular, Relational, Time Series and Forecast models; cuML, cuOpt | Predictions, anomalies, forecasts on structured data |
Palantir, Databricks and Snowflake are already building on these components (IntelligentHQ).
Where the output ends. Each function produces candidate knowledge: relevant passages, inferred relationships and statistical predictions. None of it is yet reconciled across sources, checked against policy, approved by a person, or owned by anyone. That boundary is the natural insertion point for a governed knowledge layer, and it is where the eGain AI Knowledge Hub plugs in.
The Trusted Knowledge Pipeline Reference Architecture
The eGain AI Knowledge Hub sits between NVIDIA’s Agent Knowledge Foundation and the agents, so nothing reaches an agent until it has been cleaned, reconciled, gap-checked, approved and attributed.
Read it from the bottom: enterprise data stays in place, NVIDIA finds, relates and predicts, the Knowledge Hub’s south side intercepts that candidate knowledge, and only certified knowledge leaves its north side. Smart Evaluator closes the loop from production answers back to the stage that needs fixing.
The Pipeline, Stage by Stage
Every piece of candidate knowledge from the NVIDIA stack passes six stages inside the eGain AI Knowledge Hub before an agent may use it.
- Intercept (south side). The Knowledge Hub subscribes to the outputs of Search, Analyze and Predict: retrieved passages, Auto-Ontology entities and relationships, and Kumo predictions. Source content stays where it lives, in line with eGain’s Federated Knowledge Orchestration model. Curated external sources from the Trusted Knowledge Network (for example regulatory or clinical references) join here.
- Clean and normalize. kForge transforms raw passages into AI-ready, human-readable knowledge: templated structure, taxonomy, topics and metadata. Duplicates, fragments and formatting noise are removed.
- Collate and reconcile. Related items are grouped against eGain Knowledge Graphs, seeded and refined by NVIDIA’s ontology output. Conflicts between sources are detected and routed for resolution; superseded versions are retired, not merely down-ranked.
- Find the gaps. WYCA (What Your Customers Ask) compares what customers and employees actually ask with what approved knowledge covers. Uncovered questions become authoring tasks rather than agent improvisation. Genie accelerates the authoring and review work.
- Supervise and govern. Human experts approve, edit or reject knowledge in governed workflows. Every item carries an owner, an approval record, an effective date and a review date. Predictions from Kumo models are labelled as predictions and governed by policy on how agents may use them.
- Certify and deliver (north side). Only approved, attributed knowledge is published to agents, with provenance attached to every answer. Delivery is open to any agent framework, including agents built with NVIDIA Agent Toolkit and running in OpenShell.
The loop that keeps it trusted. Smart Evaluator watches agent answers in production, classifies errors and attributes each to a stage: source, Knowledge Hub, retrieval or agent workflow. Fixes flow back into the pipeline, so quality improves with use instead of drifting.
Governance, Human Supervision and Liability
An enterprise can only defend an agent’s answer if it can show who approved the knowledge behind it, when, and under which policy. The Trusted Knowledge Pipeline makes that record a by-product of normal operation.
NVIDIA governs what an agent may do: OpenShell isolates execution and enforces dynamic privileges. eGain governs what an agent may know and say. The two controls are complementary, and a regulated enterprise needs both.
| Control | What It Guarantees | Evidence Produced |
|---|---|---|
| Ownership | Every knowledge item has a named accountable owner | Owner of record per item |
| Human approval | No knowledge reaches an agent without expert sign-off, at the level of rigor policy requires | Approval history and versions |
| Effective dating | Superseded knowledge is withdrawn on a set date, not left to rank | Effective and expiry dates |
| Source provenance | Each answer traces to approved items and their original sources | Citation chain per agent response |
| Use policy | Rules for how agents may use predictions, external sources and sensitive topics | Policy version applied per response |
| Audit and remediation | Errors are attributed to a stage and fixed at the root | Evaluator diagnostics and fix log |
Liability. Accountability for agent outcomes ultimately rests with the enterprise. The pipeline lets it accept that liability on firm ground: answers come only from knowledge it has approved, every answer is traceable, and every error has an owner and a remediation path. For compliance, risk and legal teams, that turns AI agents from an open-ended exposure into a governed business process.
Illustrative Use Cases
The pipeline matters most where a wrong answer carries regulatory, financial or clinical consequences. Three illustrative scenarios:
| Industry | NVIDIA Stack Contributes | Knowledge Hub Adds | Agent Outcome |
|---|---|---|---|
| Retail banking | Retrieval across product terms, fee schedules and procedures; Kumo flags unusual account patterns | Reconciles conflicting fee documents; maps federal and state obligations to rules of engagement; withholds unapproved language | Service agent explains a fee or dispute process with the current, approved answer and a cited source |
| Insurance | Retrieval across policy wordings and claims guidance; ontology links policy, coverage, claim and customer | Resolves wording conflicts across policy versions; routes gaps found in claims questions to underwriting experts | Claims agent gives a coverage answer tied to the customer’s actual policy version |
| Health insurance and care | Retrieval across member benefits and internal procedures; forecasts on care demand | Intersects member entitlements with curated clinical references from the Trusted Knowledge Network; clinician sign-off on guidance | Member agent answers within entitlements and approved clinical guidance, and hands off when it should |
In each case NVIDIA’s stack makes the agent fast and well-informed, and the Knowledge Hub makes it right, current and defensible.
The message to the market is direct: NVIDIA has said agents need knowledge. Enterprises need that knowledge to be trusted. Together, we deliver both.
Sources
- GTC Berlin 2026: Live Updates From NVIDIA CEO Jensen Huang’s Keynote, NVIDIA, Oct 21, 2026
- NVIDIA Introduces Agent Knowledge Foundation at GTC Berlin 2026, IntelligentHQ, Oct 21, 2026
- AI Agents Make Security A Fundamental Engineering Problem: Huang, BW Businessworld
- NVIDIA wants AI agents on a leash with enhanced security measures, Digit
