Ask a chief information officer about AI in 2023 and the conversation began with a single question: which model is best? Three years on, a striking share of new research suggests that was the wrong question all along. Enterprises are pouring money in. Some 84 percent plan to raise AI investment this year, according to the 2026 Buyer Behavior Report, and in India spending jumped 119 percent in a single year. Yet only 25 percent of leaders say generative AI has actually transformed their business. The gap between what companies spend and what they get has become the defining story of enterprise AI, and a cluster of reports out this week points to the same culprit.
The culprit is not model quality. As Tarun Dua of E2E Networks argues in the Hindustan Times, intelligence has lost the one property that made it valuable: scarcity. Two years ago capable models were rented from a handful of providers. Today open-weight systems such as Qwen, Gemma and NVIDIA's Nemotron sit within striking distance of the best proprietary ones, and the price of inference has fallen more than 280-fold for the same class of model by one Stanford estimate. When intelligence becomes something you consume like bandwidth or storage, owning the smartest model stops being an edge. What turns scarce instead is dependable work: a system that delivers the same result at nine in the morning and again at two the following night, without a person standing over it.
That is harder than it sounds. An agent that gets each step right 95 percent of the time completes a twenty-step workflow barely a third of the time. The unforgiving arithmetic explains why writers at McKinsey, the World Economic Forum and Global Banking and Finance keep circling the same unglamorous answer: value has moved to the orchestration layer, the harness of identity, permissions, routing and monitoring that decides which model handles a task and whether a human must approve the next step. ServiceNow's maturity index found only 22 percent of Indian enterprises have the testing and governance to run agents safely, and just 11 percent have reached autonomous workflows. Eagle Hill Consulting found 84 percent of leaders naming at least one cultural barrier to AI success, even though only 9 percent thought culture mattered in the first place.
There is a competitive edge buried in this, and it is not the one vendors advertise. Satya Nadella conceded in July that AI labs learn their customers' know-how from their own usage. Palantir's Alex Karp put it more bluntly, saying enterprises are paying to lose their competitive edge. The workflows a company automates are its moat, which raises an awkward question about who gets to watch them run. That logic is fuelling interest in open-weight models on infrastructure a company controls, so the corrections that would otherwise train a supplier's system stay in house.
The same squeeze is reshaping who profits. An analysis in the AI to ROI newsletter notes that among the top 35 AI startups, OpenAI and Anthropic now capture 89 percent of every dollar spent, while AI-native application companies are boxed in from every side: model makers moving up the stack, incumbents like Salesforce and ServiceNow rebuilding as AI-first, and coding agents letting enterprises simply build what they once bought. The survivors, from legal firm Harvey to clinical-documentation company Abridge, share a pattern. They start with a painful, expensive workflow, accumulate proprietary context, and sell measurable outcomes rather than access to a model. The lesson for everyone else is the same: in an era of plentiful intelligence, the advantage belongs to whoever can turn it, reliably and on their own terms, into finished work.