TL;DR: Personalized learning has been the promise of education technology for thirty years. The gap between promise and delivery isn’t a technology problem. It’s an ontological alignment problem. And the same force that’s fragmenting the enterprise application layer is now fragmenting the learning delivery layer. As AI systems become the channel through which knowledge reaches learners, the content providers that survive and thrive will be the ones whose knowledge is explicitly modeled, not just packaged. The course catalog era is ending. The ontological asset era is beginning.

In the previous piece in this series, I argued that the organizational ontology — the explicit, shared model of what a business’s objects, relationships, and logic mean — is becoming the new competitive moat as the application layer fragments. Every team building its own tools raises a question: what holds the business together underneath?

The learning domain is facing a version of the same disruption, but the implications land somewhere different. The fragmentation isn’t just happening at the enterprise end, inside organizations deploying AI tools that route learning content to workers. It’s happening at the delivery end too. The channel through which knowledge reaches a learner is no longer a portal, or a course, or even an LMS. It’s an AI system, acting in context, pulling on whatever knowledge assets it can find and use. The question for content providers, for anyone whose business is making knowledge available to people who need it, is whether their assets are structured to survive that transition.

The promise and the gap

The dream of personalized learning predates the web. By the 1980s, Intelligent Tutoring Systems were attempting to do what good human tutors do naturally: model the learner’s current understanding, identify the gap between where they are and where they need to be, and deliver exactly the right intervention at the right moment. John Anderson and his colleagues at Carnegie Mellon built tutors grounded in cognitive science (first for geometry proofs and LISP programming, eventually for algebra) that could trace a student’s problem-solving steps in real time and respond accordingly. Carnegie Learning’s Cognitive Tutor, the commercial descendant of that work, was validated in a large-scale randomized trial across 147 school sites and roughly 25,000 students. It remains one of the most rigorously evaluated educational software products ever built.

But it never generalized. Building a cognitive model of a single subject domain (one that could track student understanding at a granular level and respond adaptively) required years of expert knowledge-engineering work. The economics never scaled beyond narrow, high-stakes domains where the investment could be justified.

So “personalized learning” evolved into something more modest: sorting learners into groups, routing them to different content paths based on prior performance, surfacing recommendations based on role or completion history. A 2024 scoping review of adaptive learning in higher education found that learner performance improved in 59% of studies, meaningful evidence that the technology works at some level. But the same research was clear about what it doesn’t do: the gains are in domain knowledge acquisition, not in the deeper development of agency, judgment, or self-directed capability. The thing the Cognitive Tutor could do at depth, the scaled versions can’t quite reach.

The gap isn’t in the aspiration. It’s in the machinery underneath.

Learning is an ontological negotiation

Here’s the framing that learning researchers have understood for a long time, even when the vocabulary was different.

Every real learning moment involves a negotiation between multiple models of the world:

The teacher’s ontology: how a domain expert structures their field. What concepts exist, how they relate, what you need to understand before you can understand the next thing, what the canonical examples are. This is the model that produced the course, the textbook, the curriculum.

The learner’s ontology: their current mental model. What they already know, how they’ve organized it, where the gaps and misconceptions are, what context they’re bringing from their actual work. Two people sitting in the same course have different starting ontologies and will build different ones from the same experience.

The organizational ontology: what competence means in this specific context. “Negotiation skills” means something specific at a procurement-heavy manufacturing company that it doesn’t mean at a consulting firm. The organization’s model of what knowledge matters, and at what level, shapes what learning is actually for.

The temporal dimension: yesterday’s learner isn’t today’s learner isn’t tomorrow’s. The negotiation isn’t between static models. It’s between models that are themselves evolving. The learner’s ontology changes through the encounter with the teacher’s. What the organization needs changes as the business changes.

Real adaptive learning, the kind a skilled human tutor provides, is continuous navigation across all of these simultaneously. The tutor’s model of the domain meets the learner’s current model, inside a specific organizational context, in a specific moment, and the tutor makes a judgment about exactly what would move the learner forward right now.

No static content path does this. No recommendation algorithm does this. Not because the algorithms are bad, but because the problem is a negotiation between ontologies, and you can’t negotiate with something that isn’t explicitly represented.

The delivery layer is fragmenting

Here’s the connection that matters for content providers specifically.

The previous piece described what happens to enterprise software when AI coding tools make application development cheap: the application layer fragments. Every team builds its own tools, and what holds the business together is the shared data layer underneath: the explicit, agreed-upon model of what the business’s objects mean.

The same fragmentation is happening to learning delivery. The portal was one channel. The LMS was one channel. Both were imperfect workarounds for the real problem, but they were at least singular: one place to put content, one place learners went to find it.

That singular channel is dissolving. Josh Bersin’s February 2026 research on the $400 billion corporate training market documents the shift clearly: organizations are moving from “learning platforms” to capability ecosystems, and 74% of companies report they are already failing to keep up with their demand for new skills despite that investment. The bottleneck isn’t content volume. It’s that the content can’t reach people in the contexts where they actually need it. AI systems are being woven into the day-to-day contexts where work happens: into the tools people use, the workflows they follow, the questions they ask mid-task. A learner today might encounter relevant content through an enterprise AI assistant, a custom internal tool, a specialized coaching application, or a general-purpose model that surfaced the right resource at the right moment. Tomorrow there will be more channels, not fewer.

EdTech Digest put it starkly in early 2026: “AI, applied narrowly, only magnifies these conditions. Applied holistically, it resolves them.” The problem isn’t the AI. It’s that AI embedded inside isolated content systems “cannot deliver institutional value. Intelligence trapped inside silos simply creates faster silos.”

For a content provider, this creates a question that goes beyond distribution. It’s not “how do we reach learners on these new platforms?” It’s “are our assets structured so that AI systems can find the right piece of knowledge for the right learner in the right context, or are we delivering course-shaped objects that only make sense inside a portal?”

A course is a container. It has a title, a duration, a target audience, a completion event. An AI system routing a learner to the right resource in context needs something different: what concepts does this content cover, how do they relate to each other, what does a learner need to already understand to use this effectively, and what organizational contexts is this genuinely applicable to? That’s not course metadata. That’s a knowledge model.

The content providers whose libraries are modeled at that level will be findable, usable, and deployable across every AI-mediated channel that forms. The ones whose libraries are packaged as courses in a catalog will be largely invisible to those systems, or worse, misapplied.

What LLMs change, and why it’s an inversion

This is where the argument turns on its head, and I mean that almost literally.

The ITS researchers of the 1980s understood all of this. Their architectures (student models, domain models, pedagogical models) were explicit attempts to represent the ontologies that needed to be aligned. The technical frameworks were sophisticated. The problem was the cost of construction. Building a domain model rich enough to support genuine adaptive tutoring required subject matter experts working with knowledge engineers over extended periods. The math simply didn’t work outside narrow, well-funded contexts.

The GOFAI tradition built formal symbolic representations first, then tried to reason with them. LLMs arrived and showed that machines could achieve remarkable language capability without hand-coded ontologies, which largely ended investment in the symbolic approach for most applications.

But now the question is no longer “can a machine understand language about this domain?” The question is “can a machine act reliably on this domain, routing the right content to the right learner at the right moment, in a way that’s contextually appropriate and auditable?”

For that, you’re back to needing explicit representations. But here’s the inversion: you no longer have to build them entirely by hand.

LLM agents can now assist in the production of symbolic model representations in ways that weren’t possible before. A content provider’s existing library — courses, modules, papers, videos — can be analyzed by an LLM agent to generate a candidate domain ontology: what concepts are covered, how they relate, what prerequisites are implied, what organizational contexts they’re relevant to. A human expert reviews and corrects rather than building from scratch. The cost of ontological engineering drops dramatically.

Research presented at IEEE EDUCON 2025 demonstrates this working in practice: an LLM-assisted pipeline extracts fine-grained topics and relationships from existing lecture materials, structures them into a knowledge graph linking curriculum, domain, and learner models, and then hands off to human experts for validation. The conclusion of the researchers is direct: the system “served as an assistant, providing the human expert with an advanced starting point for the structuring of the learning content.” The knowledge engineer doesn’t disappear; the economics of knowledge engineering change fundamentally.

The negotiation between ontologies that was always the right model for adaptive learning is now computationally tractable. Not because LLMs replace the symbolic representations (they don’t), but because LLMs dramatically reduce the cost of building and maintaining them.

What this means for content providers

The fragmentation of the delivery layer changes the competitive landscape for anyone in the business of making knowledge available.

The library is an asset, but only if it’s modeled. The value of a content library in an AI-mediated world isn’t in its packaging. It’s in how explicitly the knowledge is represented and how well it maps to the contexts of the organizations and learners it serves. A library that has been ontologically modeled — concepts tagged, relationships mapped, prerequisites characterized, organizational applicability defined — is a fundamentally different asset than the same content sitting in a portal. The first can be found and deployed by any system that speaks the right standards. The second waits for a learner to navigate there.

The skills and competency bridge matters. The path from “what knowledge exists in this library” to “what this specific learner needs right now” runs through an organizational skills model: what roles exist, what competencies they require, how those competencies relate to actual knowledge. It’s worth being precise about the word: a skills taxonomy lists and organizes capabilities in a hierarchy; a skills ontology specifies how they relate to each other, which ones are adjacent, what the progression paths are, and how they connect to roles and knowledge. The distinction matters practically: taxonomies go stale; well-constructed ontologies are self-evolving. LinkedIn’s 2025 Workplace Learning Report found that 91% of talent leaders consider continuous skills development essential for business growth, yet nearly half can’t accurately identify the real skills inside their own workforce. That gap, between what organizations need and what they can see, is exactly the space the competency bridge fills. Content providers who can map their knowledge to organizational skills models will be far more useful to enterprise buyers than those who offer only self-contained courses. The bridge is the value.

The alignment layer is where new partnerships form. The hard work isn’t building any single ontology. It’s the alignment between them: between what a content provider knows and what a specific organization needs, and between what an organization needs and what a specific learner is working on today. The providers who invest in making that alignment layer explicit, and who make it legible enough for AI systems to use, will be the ones enterprises want to build with, not just buy from.

A question worth sitting with

The previous piece closed with: “What is our organizational ontology, and who controls it?”

For content providers, the parallel question is: “How explicitly is our knowledge structured, and could an AI system find the right piece of it for the right learner, in the right context, at the right moment?”

Not “are our courses well-organized.” Something harder: is the knowledge inside them represented in a way that travels outside the container? Can it be understood by systems that have never seen the course itself?

The thirty-year promise of personalized learning is finally within reach, not because the algorithms got better, but because the thing that was always missing (explicit, maintained, ontologically modeled knowledge that can meet learners wherever they are) is becoming feasible to build. The delivery channels will keep multiplying. The content that’s structured to work across all of them will be in a fundamentally different position than the content that isn’t.

First in this series: When Everyone Has Their Own Software, the Ontology Holds It Together