Here is the number the AI industry does not want you to see: fewer than 10% of enterprises that experimented with agentic AI have managed to scale it into anything that delivers real value.

That is not from some bearish analyst or AI skeptic blog. That is McKinsey, the consulting firm that has arguably done more to hype AI to boardrooms than any other organization on the planet. When McKinsey publishes a report titled "Building the Foundations for Agentic AI at Scale" and the headline number is a 90% failure-to-scale rate, you should pay attention.

The details are even more damning. Nearly two-thirds of enterprises worldwide have experimented with AI agents. That sounds impressive until you realize it means the majority of companies that tried them got stuck in the pilot phase, running demos that impressed the C-suite but never survived contact with production systems, real data, and actual business processes.

A separate McKinsey survey found that 23% of organizations claim they are scaling agents in at least one function. But high performers, the companies actually getting results, are roughly three times more likely than their peers to be scaling across the enterprise. Translation: most companies scaling in "at least one function" are running a single chatbot that handles tier-one support tickets and calling it enterprise AI.

The reasons are not mysterious. Deploying agents at scale requires policy frameworks, retrieval systems, audit trails, and governance structures that most companies have not built. It requires clean data, which most companies do not have. It requires cross-functional alignment, which most companies actively resist. And it requires someone to own the problem end-to-end, which in most organizations means nobody owns it.

Gartner is projecting that 40% of enterprise software will feature task-specific AI agents by the end of 2026. That prediction will almost certainly come true, because embedding an agent in software is easy. Getting the agent to do something useful in a messy corporate environment is the hard part, and it is the part that 90% of companies are failing at.

Meanwhile, 89% of enterprises say they plan to increase their AI investments in 2026 and beyond. Think about that. Nine out of ten companies plan to spend more on a technology that nine out of ten companies have failed to scale. That is either extraordinary optimism or the sunk cost fallacy operating at industrial scale.

There is a bright spot. Banks implementing agentic AI for KYC and anti-money-laundering workflows are reportedly seeing 200% to 2,000% productivity gains. But banking compliance is a near-perfect use case: highly structured data, clearly defined rules, massive volumes of repetitive decisions, and enormous regulatory pressure to get it right. Most business processes do not look like that.

The agentic AI moment feels a lot like the cloud migration moment of 2015. Everyone agreed it was the future. Everyone started pilots. And it took most companies five to seven years to figure out how to actually do it well. The difference is that AI vendors are burning through cash at a pace that makes 2015 cloud spending look quaint. OpenAI alone is on track to burn $85 billion this year.

The companies that will win the agentic AI race are not the ones buying the most tokens. They are the ones doing the boring work of data transformation, governance, and process redesign that makes agents actually useful. McKinsey knows this. The question is whether its clients are listening.