The agents arrived before the operating model

Two Gartner forecasts frame the year we are in. In August 2025, Gartner predicted that 40 percent of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5 percent in 2025. Two months earlier, the same firm predicted that over 40 percent of agentic AI projects will be canceled by the end of 2027, citing escalating costs, unclear business value and inadequate risk controls. Microsoft's 2026 Work Trend Index adds the usage curve: active agents in Microsoft 365 grew 15 times year over year.

So the agents are coming either way; they are being shipped into the software your teams already use. What gets canceled is not the technology. It is the projects: the human wrapper of budgets, expectations and accountability around the technology. The software arrived before the operating model.

Agents fail the way unmanaged hires fail

I have watched this closely for a decade, first automating financial back offices, now running my own companies on agents daily. The failure pattern is rarely the model. It is that the agent has no role definition, no manager, no feedback loop and no offboarding. Imagine hiring a capable junior, giving them no job description, no supervisor and no probation review, and then concluding from the mess that juniors do not work.

Gartner's cancellation note names the same disease from the vendor side: agent washing, existing products rebranded as agentic. Gartner estimates only about 130 of the thousands of self-described agentic vendors are real. If a company cannot say what an agent owns, it also cannot tell a real one from a rebadged macro.

The boring fix

• Write a job description per agent: scope, inputs, outputs, what it must never touch.
• Give every agent one named human owner. Shared ownership is no ownership.
• Define escalation before deployment: which decisions come back to a person, always.
• Review output by default. The Work Trend Index found 86 percent of workers treat AI output as a starting point that needs human oversight. That is not a weakness of the tools; it is the new shape of the work.
• Keep a kill rule. An agent that cannot show its value in a quarter gets retired like any failed process, without sentimentality.

In the Framework's terms: machines can only take over process that has been made explicit. The demo-to-production gap where agentic projects die is almost always a legibility gap. Companies that never wrote their processes down are now discovering they cannot delegate what they cannot describe.

The winners of the next two years will not be the companies with the most agents. They will be the ones that can run a mixed roster of people and agents without drama: clear roles, clear owners, clear reviews. That is not an AI capability. It is a management capability, and it is trainable.