Ask most executives what will give them an edge in artificial intelligence and they will point at the model: which lab, which version, how many parameters. A run of new industry reports this week suggests they are looking in the wrong place. After studying more than 100 of its own internal AI transformations, Microsoft published a playbook with a deliberately deflating message. The winning enterprise architecture, it argues, puts proprietary evaluations, context and learning loops above the foundation model itself, because the models are increasingly interchangeable and the context is not.
The point lands harder when several unrelated voices arrive at it independently. Writing in CIO, one practitioner describes watching companies stash their business knowledge in GitHub repositories, prompt libraries, shared folders and one-off retrieval pipelines, then lose track of it the moment finance changes how it defines revenue or legal rewrites a policy. His prescription is a "context development lifecycle," borrowing the discipline software engineering built for code: someone owns each definition, changes are reviewed and tested, and every version has an effective date and a history. A separate analysis frames the same gap as a knowledge problem, warning that an AI system built on undocumented expertise will not just be imprecise, it will be "confidently imprecise," which is the more dangerous failure. When the expert who was the only person who understood a workflow retires, the AI inherits the hole they left.
None of this is glamorous, which is precisely why it gets skipped. A twelve-point framework laid out by Anupam Arora of India's Bharat Forge, drawn from real deployments on a factory floor, keeps returning to the same unfashionable foundations: clean, well-owned data is the real moat, because top-tier talent and the latest chips can be bought and proprietary knowledge cannot. He points to General Electric's Predix platform, a multibillion-dollar bet that collapsed trying to scale before it had proved value, as the cautionary tale everyone quotes and few heed.
There is a tension worth sitting with here. The reports lean on some eye-catching statistics, an MIT finding that 95 percent of organisations saw no measurable profit impact from generative AI, a widely cited claim that only 6 percent capture real value, that are easy to brandish and hard to verify precisely. But the underlying argument does not depend on any single number. If the models are converging and available to everyone, then the differentiator has to be what you feed them and how you govern it. That reframes the AI project from a shopping decision into an unglamorous, years-long discipline of writing down what your company actually knows. The firms treating it as procurement will keep buying newer models and wondering why the returns never show up. The ones treating it as knowledge management may find the moat was theirs all along, if only they had documented it.