Unstructured Data Is the New Moat: Why Context Beats Model Size

Every few months a new frontier model tops the leaderboards, and every few months the gap between the top models narrows. For businesses, the strategic implication is easy to miss: if everyone can rent roughly the same intelligence, the intelligence itself stops being the differentiator.

What cannot be rented is your context. The overwhelming majority of a company’s knowledge, by most industry estimates, around nine-tenths of it, is unstructured: contracts, proposals, email threads, meeting notes, SOPs, support histories. It is the record of how your business actually works, and it exists nowhere else on earth. Aaron Levie of Box has spent years making a version of this argument: the companies that win with AI will be the ones that get their unstructured content organised, governed and connected to the models, because that content is precisely what a foundation model does not know.

Generic model, generic answers

Point a frontier model at a question about your pricing history, your contract terms or your delivery exceptions and, without access to your data, it produces something fluent and useless, or worse, fluent and wrong. The model is not broken. It is starved. Its answers can only ever be as specific as the context it is given.

This is why we treat the context layer as the core asset in every engagement. Mapping your clients, projects, obligations and workflows into an explicit, connected structure is what turns a rented model into a system that answers like a fifteen-year employee. The model can be swapped every quarter as prices fall. The context layer appreciates.

The governance half of the story

There is a second half to the argument that gets less airtime: connecting AI to everything is dangerous if permissions and provenance do not travel with the data. An agent that can read every contract must still respect who is allowed to see which contract, and every answer should be traceable to its source. Governance is not the tax on the moat, it is part of the moat, because it is what makes the system safe enough to wire into real operations.

Where to start

Not with a model. Start with an inventory: which decisions do your people make repeatedly, and which documents do they reach for to make them? That intersection, high-frequency decisions, retrievable context, is the first candidate for a Digital Brain. Organise that slice, connect an agent to it, measure task resolution, then widen. The moat is dug one workflow at a time, and unlike a model subscription, you own it when it’s done.