On April 28, Gartner published from its Digital Workplace Summit in London an uncomfortable projection for the entire sector: in 2028, an average Fortune 500 company will have more than 150,000 agents artificial intelligence in use, down from less than 15 in 2025. It is a multiplication of more than ten thousand times in three years. And only 13% of organizations believe they have adequate governance to manage them.

The number arrived just when The Wall Street Journal revealed the other side of the phenomenon. coworkthe tool Anthropic which allows any employee without programming knowledge to create their own artificial intelligence agents to automate tasks, is one of the platforms that triggered the problem.

Companies like Lyft, DaVita, GitLab, FICO and Magnum Ice Cream, owner of Ben & Jerry’s, acknowledge that they have already lost control over how many agents run within their organizations.

For two years, AI vendors pushed a seductive idea: any employee, without knowing how to program, would be able to create their own agents. Platforms like Claude Cowork from Anthropic and open source tools like OpenClawlaunched in November 2025 and with more than 100,000 stars on GitHub, the platform where programmers publish and share software, made that promise an operational reality.

The result: DaVita, the American kidney dialysis chain, has more than 10,000 agents created by your employeesas stated by its technology director Madhu Narasimhan to the WSJ. FICO, the company behind the most used credit score in the United States, adds dozens of new agents every day among its 3,500 employees. Mike Trkay, the company’s chief technology officer, described it unfiltered: Agents are created “literally every day, and at almost every level of the hierarchical structure.”

What was sold as democratization of access to AI ended up becoming a classic problem for technology departments: no one knows how many agents there are, where they run, what data they touch, or how much it costs.

Michael Friedlanderchief technology officer for the Americas at Magnum Ice Cream, put it in direct terms: “There are going to be tokens, then there are going to be costs, and then you end up with: how do we handle this so that it is financially responsible?”.

Every time an agent works, it consumes tokens, the computing units that artificial intelligence companies charge for each operation. The tokens are paid for by the company. And no one measured that expense before the massive implementation.

DaVita built an internal platform to cut back on low-performing agents, that is, reduce the processing work assigned to them, and increase spending on high-performing ones. It is the clearest admission that The problem is no longer technical but financial.

15 years ago, technology managers were fighting against employees who installed Dropbox without permission. The industry called this shadow technology: software that enters the company outside of all control. What we have now is that same shadow technology on steroids: agents execute actions, connect with other programs, access databases and make decisions at machine speed.

Microsoft published in its Cyber ​​Pulse 2026 that the 80% of the Fortune 500 already operate active agents, but only 10% have a formal strategy to manage them. The report documents specific cases of companies that discovered dozens of autonomous agents running in business units that they had never approved or inventoried.

Max Gossthe senior Gartner analyst who signed the statement, warned against the regulatory temptation of blocking. “If employees can’t work with authorized tools, they are going to bypass organizational controls and use shadow AI, which presents much greater risks,” he said in London, referring to the use of artificial intelligence outside of company-approved channels.

Anthropic and the other agent creation platform vendors are in an awkward position: They sold ease of use as a competitive advantage, and that same ease is now the problem their corporate clients don’t know how to manage.

The company announced new features for technology administrators: expense tracking, usage analytics, audit logs, curated libraries of add-ons. They are late to a phenomenon that is already off scale.

The question that matters is not whether companies will have 150,000 agents in 2028. The question is who is going to pay the token bill, who is going to audit them when something goes wrong, and what happens when an agent with write permissions deletes a production databaseas already happened in Replit and documented Microsoft’s Cyber ​​Pulse.

The promise of democratization has taken its toll: it is paid by the technology department, silently, while trying to inventory what their own suppliers pushed them to deploy without control.