Read the people paying closest attention, and the answer is the same. Top global VC fund A16Z (Andreessen Horowitz) published a piece on it this week (Yoko Li, "Knowing When to Stop").
Palantir Technologies has built an empire on it for the better part of a decade. Gartner has been saying for years that the majority of enterprise AI never succeeds, and that the reason is usually the data, not the model.
It really is quite simple.
The models were never the problem. LLM models are extraordinary, and they get better every quarter regardless of what any single company does.
The only thing that has actually held enterprise AI back is how these models interact with data. Data in the right structure, in the right order, so the model has something coherent to work in.
That structure has a name. And that name is ontology.
In simple English, an ontology is just a map of your world that a machine can reason over. What things exist, how they connect, what rules hold them together. Your assets, your counterparties, your contracts, the filing each number came from, all as defined, connected objects rather than ten thousand PDFs nobody can query. That is the idea Palantir turned into a category: the model is a commodity, the ontology is the moat.
The a16z piece explains the mechanism underneath it. An AI system has no internal sense of when it is finished. "Done" is not a property of the work; it is a judgment produced by the system around it. A model only converges on the right answer when it has an organised target to measure itself against.
Give it structured data, and it works. Take it away, and it generates confident nonsense until the sentence sounds complete.
From my experience, I believe people will be shocked at how much even the older AI models can achieve (not just in accuracy, but in reduced token use/cost) when working with structured, updated data.
AI does not need to be smarter to change your business. It needs the right conditions to work in: not the raw, unstructured chaos where it hallucinates, and not something so rigid it cannot move, but the Goldilocks state in between, where the data is organised, validated, defined and traceable, so the model finally has what it needs to perform optimally and consistently. Oh, and materially cheaper (8-20x on mining technical reports, to be precise).
That is the actual work, and it is unglamorous, which is exactly why so few do it, and so many keep waiting for the next model to save them. It won't. The next model will be brilliant and it will still not know whether your reserve figure is right, because that was never a question about the model. It was always a question about the data underneath it.
Get the ontology right, and enterprise AI stops being a science experiment.
It really is "that simple".
Less searching. More strategising.™
Where does your team's data infrastructure sit today?
Answer 10 questions. Get a private diagnostic on your AI readiness, in minutes.
Less Searching. More Strategising.™
See the platform running on real mining data. Book a demo to see what this looks like for your team.