AI Agent Proliferation Spirals Out of Control at Fortune 500 Companies

In the span of three years, the Fortune 500 is about to undergo a transformation that no one quite planned for. Today, a typical large corporation runs fewer than fifteen artificial intelligence agents. By 2028, that number will balloon to over 150,000—a multiplication so staggering it reads almost like a typo. Yet when Gartner released this projection from its Digital Workplace Summit in London on April 28th, it landed alongside a more troubling statistic: only thirteen percent of organizations believe they have the governance structures in place to manage what's coming.

The proliferation has a name and a face. Anthropic's Cowork platform, along with open-source tools like OpenClaw, promised something seductive and seemingly harmless. Any employee, regardless of technical skill, could now build their own AI agents to automate routine work. No coding required. No gatekeepers. No waiting for the IT department. It was democratization in its purest form—and it worked exactly as advertised. The problem is that it worked too well.

DaVita, the American dialysis chain, discovered it now hosts more than 10,000 agents created by its own workforce, according to the company's chief technology officer Madhu Narasimhan. At FICO, the firm behind the credit score most Americans carry, new agents spawn daily across its 3,500 employees. The company's technology chief Mike Trkay described the situation without euphemism: agents are being created "literally every day, and at almost every level of the organizational hierarchy." Companies like Lyft, GitLab, and Magnum Ice Cream—owner of Ben & Jerry's—have all reached the same uncomfortable realization: they no longer know how many agents are running inside their own walls.

What was sold as liberation arrived as a management crisis. The agents don't just sit idle. They execute actions. They connect to other systems. They access databases. They make decisions at machine speed. And each one consumes tokens—the computational units that AI companies charge for with every operation. The bill arrives at the company's door, but no one measured the expense before the deployment began. Michael Friedlander, technology director for Magnum Ice Cream, laid out the sequence plainly: "There will be tokens, then there will be cost, and then you end up with: how do we manage this to be financially responsible?" DaVita built an internal platform to throttle underperforming agents and redirect spending toward the ones that deliver results—a tacit admission that the problem had shifted from technical to financial.

This mirrors a crisis the technology industry thought it had solved fifteen years ago. Back then, IT departments battled employees who installed Dropbox without permission—what the industry called shadow IT. What's happening now is shadow IT on steroids. Microsoft's Cyber Pulse 2026 report found that eighty percent of Fortune 500 companies already operate active agents, but only ten percent have a formal strategy to govern them. The report documents concrete cases of companies discovering dozens of autonomous agents running in business units that never approved or even knew about them. When an agent with write permissions deletes a production database—as already happened at Replit—the question of accountability becomes urgent and unanswerable.

The vendors face an uncomfortable position. They sold ease of use as a competitive advantage, and that same ease is now the problem their corporate customers cannot manage. Anthropic has announced new features for technology administrators: spending controls, usage analytics, audit logs, curated plugin libraries. They arrive late to a phenomenon already spiraling beyond scale. Max Goss, the senior Gartner analyst who signed the projection, warned against the temptation to simply lock down access. "If employees can't work with authorized tools, they'll bypass organizational controls and use shadow AI, which presents much greater risks," he said in London.

The real question isn't whether companies will have 150,000 agents by 2028. They probably will. The question is who pays the token bill, who audits them when something breaks, and what happens when governance fails at scale. The promise of democratization has collected its cost. The technology department pays it, quietly, while trying to inventory what its own vendors pushed it to deploy without control.