A knowledge-graph-grounded counter-argument to Jaya Gupta's LinkedIn article: enterprises that separate model reasoning from model learning can leverage frontier AI without surrendering their institutional know-how.
Jaya Gupta’s thesis: AI creates enormous value but enterprises struggle to capture it. When many firms in the same industry — insurance carriers, biotechs, manufacturers — run their workflows through the same shared frontier model, each firm’s hard-won judgment (fraud patterns, underwriting exceptions, pricing intuition) gets absorbed as training signal. Data-retention and no-train clauses protect the input, not the training signal. The result: the first adopter banks a real productivity gain, but the durable value — the compounding learning curve — accrues to the model vendor, not the enterprise.
Kingsley Uyi Idehen’s comment, as it appeared on the source thread.
Compelling insights, but I don’t think this outcome is inevitable.
The examples cited—claims, approvals, fraud signals, workflows, and expert judgment—can all be represented in enterprise knowledge graphs as machine-computable entities, relationships, and inference rules.
When those knowledge graphs are exposed using Linked Data principles and protected with fine-grained, attribute-based access controls (ABAC), a frontier model can reason over enterprise knowledge without acquiring or retaining it as part of its own parameters.
The real issue isn’t whether AI can reason over institutional knowledge. It’s whether the underlying architecture allows that knowledge to become part of the model’s learning curve.
In other words, the choice isn’t between using frontier models and protecting enterprise know-how. With the right technology and open standards, you can do both. The real choice is between architectures that conflate reasoning with learning and architectures that deliberately separate them, keeping enterprise knowledge protected while still making it available for authorized AI use.
This is the detailed follow-up to Kingsley’s LinkedIn comment that LinkedIn’s comment-length limits couldn’t hold — a quote from the article, a verdict, then the argument.
These are all entities in an enterprise knowledge graph.
These are relationships (or, more generally, attributes — a superset of properties) connecting those entities to other entities in an enterprise knowledge graph.
An enterprise knowledge graph comprising this data, information, and knowledge becomes an artifact protectable via fine-grained attribute-based access control (ABAC) combined with hyperlinks for entity and relationship denotation. A frontier model sees a semantic web constructed from hyperlinks but cannot necessarily dereference what they denote — so it can reason over enterprise knowledge without acquiring or retaining it in its own parameters. Knowledge capture and knowledge exfiltration are fundamentally different problems; the goal is to prevent enterprise knowledge from becoming part of the model itself, not to prevent reasoning over it.
Left unconstrained, an agent harness could reconstruct its own persistent ontology of entities, relationships, workflows, and inference rules from what it observes across sessions — turning every interaction into training signal by another name. That risk closes only when those workflows are mediated through enterprise knowledge graphs protected by fine-grained ABAC: the harness can query and reason over authorized entities to complete a task, but has no channel to retain, enrich, or carry forward what it learned once the session ends. The enterprise retains ownership of its learning curve; the model gets task completion, not a standing internal map of how the enterprise operates.
Again, not if those organizations construct enterprise knowledge graphs using Linked Data principles — effectively manifesting a Semantic Web over their internal data spaces (databases, knowledge bases, file systems, and APIs) — and protect those resources using fine-grained ABAC. In that architecture, the frontier model contributes reasoning capability while the enterprise retains control over its institutional knowledge. The model can follow hyperlinks, dereference only those resources authorized for the requesting user and agent, and execute approved workflows. This separation between a model's own probabilistic reasoning and the enterprise's deterministic control over reasoning and learning — enforced through ABAC access control lists (ACLs) — is precisely what prevents institutional know-how from becoming part of a shared model’s learning curve. The article’s concerns become valid only when organizations adopt architectures that allow enterprise knowledge to leak into shared model training — the prevailing, but not inevitable, practice today.
The choice is not between using frontier models and protecting enterprise know-how. It is between architectures that conflate reasoning with learning and architectures that deliberately separate them using ABAC ACLs. Enterprises that represent institutional knowledge as machine-computable knowledge graphs, expose that knowledge through Linked Data principles, and enforce fine-grained ABAC can leverage frontier AI while ensuring that their institutional knowledge remains their own.
The lightweight ontology behind this response models two architectures for deploying frontier AI against enterprise knowledge.
High capture risk, no enforced control — versus low capture risk, fine-grained ABAC + hyperlink-based authorized dereference.
Interactive view of the ontology itself — the two subclasses, the two contrast properties, and the instance each is bound to. Drag to reposition, click a node or edge label to open its full description.
A machine-computable representation of an organization's data, information, and knowledge as entities, relationships/attributes, and inference rules, denoted by hyperlinks.
Fine-grained access control that grants or denies dereference/use of a resource based on attributes of the requester, resource, action, and context, rather than fixed roles alone.
The use of Internationalized Resource Identifiers (hyperlinks) to name and denote entities and relationships within a knowledge graph, independent of whether the referent is dereferenceable by a given requester.
The act of resolving/looking up an IRI to retrieve the resource or data it denotes; access to dereferencing can be authorization-gated independently of the ability to see the IRI itself.
The distinction between a model reasoning over enterprise knowledge to complete a task (capture/use) and that knowledge becoming incorporated into the model's own trained parameters (exfiltration).
The orchestration layer (tools, memory, workflow control) through which a frontier model interacts with enterprise systems and data spaces to execute tasks.
The condition in which an organization's cumulative operational learning (mistakes, overrides, judgment) is retained and improved within the enterprise's own systems rather than becoming shared-model training signal.
Use of IRIs to identify things, HTTP IRIs so they can be looked up, provision of useful RDF information upon lookup, and inclusion of links to other IRIs for discovery — applied here across an enterprise's internal data spaces.
A web of machine-readable, linked, authorization-gated data spanning an organization's databases, knowledge bases, and file systems, built on Linked Data principles.
An architectural principle distinguishing a model's use of enterprise knowledge to reason and complete tasks from that knowledge being retained as training signal for future model versions.
A sampling of the surrounding comment thread, for context on how the argument landed with others.
Intersting perspectives. What stood out for me is that this feels much bigger than enterprise AI adoption. It raises a broader question about what organizations should consider their true source of competitive advantage as AI becomes part of everyday work. I tried to capture how I thought about it in the attached map. Curious how others are thinking about this shift.
The risk here is tied to how much of your workflows are probabilistic vs deterministic. A risk that can be further mitigated by governing the context packet your substrate hands over to any model. So many have been promoting ‘context’ vendors — lining people right up to the very thing you warn of here. Sending an agent rampant through your systems to help you ‘find’ something is a feast of context, worse if you have no data/IP hygiene. I think about my time as an auditor, when someone didn’t feel like looking for a file so they gave me access to the folder to find it myself. When your house is in order, it helps us both. When it’s not, I learn far more than you probably would have liked me to.
If you don’t own the friction, you’re not the business. You’re the training data.
Now do they really do this kind of research? Can they do it all at once? Is a company value creation really a model/an equation? My intuition is that the singularity of individuals’ impact has more weight than common business culture admits. The common counter argument is saying that Outlook/Gmail has been having the last 30 years of company confidential and decisions’ data completely available, yet did they use it in any meaningful business way? Anyhow, local and open source will definitely pick up.
Interactive force-directed view of every entity and relationship in the companion RDF. Click a node or edge label to open its full description via URIBurner.
This page was generated from a companion RDF-Turtle knowledge graph (schema.org vocabulary plus a lightweight architecture-contrast ontology) modeling Kingsley Idehen’s response to Jaya Gupta’s July 9, 2026 LinkedIn article on AI value capture, then rendering that graph as this interactive infographic and companion Markdown document.
The mistake is thinking the moat is the data. The moat is usually the taste. In insurance, pharma, trading, procurement, law, medicine, the edge is not just what you know. It is how your best people decide when the facts are incomplete. Shared AI turns that private judgment into public infrastructure. So the question for CEOs is not “does the vendor train on our data?” It is: are we teaching the industry how to think like us?
Also flagged an AI-generated explainer video of the article, embedded below.