Thirty-Two Percent of Enterprises Stopped Buying Software. That's the Best Possible News for Vertical AI.

Thirty-two percent. That's the share of large organizations — those with more than a billion dollars in revenue — that skipped at least one software purchase in 2026 because their teams could build the equivalent using AI agents. McKinsey's State of AI 2026 puts a number on the shift that has been visible anecdotally but lacked scale-level evidence until now.

The naive read is that enterprise software is dying. It isn't. What's dying is software that automates generic, well-understood workflows — the kind a competent team with Claude Code or Cursor can recreate in days. That category is large. It includes dozens of point solutions that raised venture rounds between 2015 and 2022 on the thesis that a specific workflow needed a dedicated SaaS vendor. It doesn't, anymore.

The durable category looks different. Forty percent of organizations with more than $1 billion in revenue are actively scaling AI agents, up from 27% a year earlier. But those deployments concentrate in functions where the training data doesn't come off the shelf: credit underwriting, fraud detection, compliance monitoring. These workflows require years of proprietary transaction history, fine-tuned evaluation infrastructure, and domain knowledge that no AI coding tool can generate from scratch. They are the antithesis of what gets commoditized.

This is the structural argument for Brazil's fintech application layer. Nubank's nuFormer credit model was trained on 139 million customer behavioral records — a dataset no competitor can replicate, and no AI coding tool can substitute. QI Tech's vehicle credit infrastructure required more than 80 business relationships across 1,000 dealership locations before a single loan was priced. The data moat isn't an accident of market timing. It's the specific asset that survives the build-versus-buy inversion.

McKinsey's data also shows a split in enterprise AI maturity that's compounding in real time. Fifty-four percent of large enterprises report scaling AI across the entire organization. The 46% that aren't will eventually face the same competitive pressure — but they will arrive later, with less proprietary training data and fewer options for catching up. The gap grows with each month of data their better-positioned competitors accumulate.

For venture capital, this resolves a debate that has run since the first wave of AI-native startups: does the application layer have durable margins, or do cheap models and free AI coding tools commoditize it away? The answer is: it depends entirely on whether the workflow requires data that only exists inside the company running it. If it does, the moat grows with every new customer. If it doesn't, you're building a feature, not a business.

The 32% who stopped buying one piece of software will eventually stop buying more. What they can't build — ever — is the proprietary behavioral dataset of a company that started collecting it five years earlier. That's the question every AI startup should be answering before its next pitch meeting.