Accumulated context, not model capability, is becoming the real moat in enterprise AI

Mike Mignano makes the argument plainly. “Context is extremely valuable,” he says. “If Granola gets inside of your organization and everyone in the organization starts using it and producing and accumulating all of this amazingly rich and valuable context, that’s not something you as an enterprise want to give up.” That single observation reframes what enterprise AI competition is actually about. It is not a race to the most capable model. It is a race to embed deeply enough that leaving becomes costlier than staying.

Ali Ghodsi, chief executive of Databricks, reaches the same conclusion from a different direction. His position is that current models are already sufficiently capable: moving from 60 to 70 percent performance on standard benchmarks is not what enterprises need. The bottleneck, in his view, is the organisational layer that tells a model what the business actually means by its own terminology, its own rules, its own definitions of a correct action. That layer does not ship with any model. It accumulates through use, slowly, inside a specific tool.

The infrastructure argument reinforces this. Julien Bek points to what happens once an enterprise has built inside a particular database: data gravity sets in, enterprise controls are configured, and the switching cost becomes structural rather than financial. His framing applies directly to AI tooling. The moment an organisation’s institutional knowledge is woven into a platform’s memory and agent logic, the platform has something that no competitor can replicate on day one, regardless of how the benchmarks compare.

Ian Silber observes that the tools to formalise this lock-in are still emerging. Durable, reusable workflows that persist across sessions are nascent, he notes, with the current experience often feeling like starting from scratch each time. When those persistent workflows arrive at scale, the switching cost calculus will harden further. An enterprise that has spent months building compounding, session-persistent context inside one platform faces a very different migration decision than one that has been working with stateless prompts.

The pricing dynamics that follow from deep lock-in are already visible in adjacent software markets. Jason Calacanis describes the private equity software playbook in terms that strip away any ambiguity: the target customer is not one who needs to be sold on the product, but one who hates a price increase yet cannot leave. Bending Spoons, as described by Harry Stebbings, illustrates the arithmetic. Tripling prices on Airtable may cause some customers to finally spend the time to migrate, but losing a portion of the base while sharply raising revenue on the remainder is, in Stebbings’s words, “a good deal.” The same logic extends to any AI platform that has become the repository of an organisation’s institutional knowledge.

There is a further wrinkle that few enterprise buyers have fully absorbed. Stebbings describes how his team’s agents have standardised on a single tool because, as he puts it, it is not worth arguing with the agents. When AI agents themselves develop stable product preferences and those preferences begin to drive organisational procurement, the switching cost acquires a new dimension. It is no longer just about what humans have built inside a platform. It is about what the agents have learned to rely on, and retraining that preference requires effort that compounds with time.

The evidence from Mignano, Ghodsi, Bek, and Silber points in the same direction. The enterprise AI vendors that win durable market positions will not necessarily be those with the most capable underlying models. They will be those that became the place where an organisation’s context lives, and made leaving that context behind an option no rational buyer will take.