The No-Moat Myth

Everyone keeps saying there is no moat, that general AI is eating everything. But there will always be opportunities to build moats. I am talking here specifically about technical moats, particularly the kind that develops when a company gets really good at one specific problem.

Companies can base a moat on owning proprietary data on how their users use their platform - but knowing how people use a platform is only useful and moat-creating when it is applied to active outcome improvement. With general AI encroaching more and more on things like memory and context, the moat is not simply whether you hold your company's context and can reason with it. It is how deeply you understand not just the brand, but the brand in the context of a category, market and space, based on thousands of relevant actions and their outcomes. Having general or even brand-specific context is not the same as having accumulated expertise in a particular domain.

General AI is general because it learns across broad sets of knowledge and tasks, with the aim of improving general reasoning. It can handle context too and hold user memory - but there is a distinction between specificity at the user level, where you understand one client's circumstances and particular problem, and specialism at the domain level, where you learn deeply within one area across many relevant cases.

The combination is powerful: interpreting a client's specific context through accumulated domain expertise. Where that involves learning across clients, it needs to be permissioned and sufficiently generalised so that IP and privacy is safeguarded, but this is the foundational premise of much specialist AI, that the category network effect will in turn fuel better outcomes over the network.

General reasoning is also not the same as proprietary learning and decision methods. How much weight should something carry? Which factors matter in this particular use case? What errors in reasoning lead to suboptimal outcomes, and what should change as a result? Those are questions that can inform updates to models and processes around a specific problem. A general model does not automatically possess a company's exclusive feedback, outcome history or methods for making those decisions. It can use agents to connect to more context but it relies on those connectors doing their job.

So today's moat is applying proprietary data and processes relevant to one area and one point in a workflow, and getting really great at that. It is not just creating something that performs better than general AI with context inputs or surface-level reasoning deployed through prompts. It is how that system develops, reacts and evolves through learning, and whether the improvement it produces is difficult for someone else to reproduce.

And that is just the technical moat. Operational advantages, distribution and network effects can create moats too. General AI becoming more capable does not mean every source of defensibility disappears.