The dominant story of the past three years was that LLMs ate structure. Feed a model enough text, the narrative went, and you don’t need ontologies, schemas, or graphs.
The model is the knowledge layer.
That story is already aging badly.
What’s actually happening is more interesting: models are commoditizing faster than anyone expected, and the moat is moving underneath them.
It’s moving to the substrate to the typed, temporal, resolved, grounded, proprietary graph that tells a system what’s actually true about your domain right now. The companies that figure this out in the next twenty-four months will define the next decade.
Here’s the mechanic. A frontier model today can do almost anything with context.
The bottleneck is not reasoning it’s assembling the right context.
And the right context, for any application that matters, is not a pile of documents. It’s a structured view of entities, relationships, provenance, and recency, composed on demand.
That is knowledge engineering. It just doesn’t look like the knowledge engineering of 2015.
The old version was humans writing triples. The new version is pipelines: extraction models parse the world at scale, resolution layers collapse duplicates, graphs hold the typed skeleton, vectors handle the fuzzy edges, and an LLM sits on top doing last-mile synthesis. Every layer matters. Get entity resolution wrong and your whole system lies fluently. Get your ontology too rigid and extraction snaps. Get provenance wrong and you can’t explain a single answer.
Operators who are paying attention have already internalized this. The interesting AI companies of 2026 are not the ones with the best prompts. They’re the ones with proprietary graphs that no one else can build because the graph was assembled from distribution they already own, signals they already have, or workflows they already sit inside. Bloomberg and Palantir spent decades building graphs before anyone called it that, and those graphs are now worth more than most AI labs.
The next wave is being built right now, in verticals where the network is the product: media, venture, healthcare, legal, sales, recruiting.
A few things follow from this that are worth front-running.
First, agents are a knowledge engineering story, not a model story.
An agent without a reliable knowledge substrate is a confident hallucinator on a longer leash. Every serious agent deployment in the next two years will be preceded by a substrate build. The teams that ship agents first will be the teams that already have the graph.
Second, data acquisition is becoming the new compute. The asymmetry is shifting from who can afford the GPUs to who has the rights to the raw material and the pipelines to turn it into structure.
Proprietary data you can legally enrich is about to get repriced.
Third, vertical eats horizontal. Horizontal AI tools keep getting squeezed because the models underneath them keep getting better and cheaper. Vertical tools with deep graphs get more defensible with every model upgrade, because the graph compounds while the model commoditizes. Every frontier release makes a thin wrapper weaker and a deep graph stronger.
Fourth, the Chief Knowledge Officer is coming back not as a title, necessarily, but as a role.
Someone on the leadership team has to own the question what does our system actually know, and how do we keep it honest. That is a strategic function now, the way data science was in 2014 and ML engineering was in 2020. Companies that treat it as infra will lose to companies that treat it as product.
The punch line: AI did not kill structure.
It made structure cheap enough to be everywhere, and valuable enough to fight over. The operators who win the next decade will not be the ones with the cleverest prompts or even the best models. They’ll be the ones who understood, early …
The graph is the product. Everything else is a detail.