Gartner: Corporate policies will not stop an agent. By 2029, 70 percent of organizations are expected to face a serious incident

On September 24, Gartner highlighted the difference between governing traditional AI and autonomous agents. Corporate policies, approval committees and documentation can define what a system is supposed to do. They cannot physically prevent an agent from carrying out a faulty operation during execution. An agent differs from a conventional chat tool because it can execute multi-step workflows, change the state of systems and act at machine speed. This, according to Gartner, is where a gap emerges between governance on paper and control at the moment of execution. Autonomous workflows use probabilistic decision-making. They can therefore cause an operational outage, make an unauthorized purchase, consume excessive amounts of paid model units or use other enterprise resources without sufficient control. Gartner predicts that by 2029 at least 70 percent of organizations with autonomous AI deployed in production infrastructure and operations will experience a serious service, security or cost incident to which inadequate runtime controls partly contribute. Gartner’s George Spafford summarizes the problem by saying that “written corporate policies cannot physically stop an agent from making a destructive mistake.” The result is a change in control architecture. It is not enough to validate the model before deployment and issue rules to employees. The company needs technical boundaries that apply to every operation: which systems the agent may open, which data it may read, how much money it may spend, how much computing capacity it may consume and when it must stop or hand the task to a human. This issue is emerging just as agents gain access to financial systems, infrastructure, software development or customer accounts. A faulty answer from a chat tool can damage one document. A faulty agent can apply the same logic across thousands of operations before anyone notices. The speed that creates economic value can therefore multiply potential harm at the same time. A day earlier, Gartner warned about the broader security consequences of uncontrolled agent proliferation. By 2028, abuse of unmanaged AI agents is expected to contribute to a quarter of enterprise security breaches. Analysts therefore point to zero-trust controls and kill switches assigned directly to individual agents. Consumption matters for financial management as well. An autonomous system can trigger many model calls or external services without an employee approving each operation. Costs can therefore escape control in the same way as security permissions. A budget limit must become a technical rule, not merely a monthly invoice review. Each agent should therefore have its own digital identity, tightly limited permissions, consumption limits and an audit record. There must also be a technical capability to stop its activity immediately. The business conclusion is uncomfortable but practical: the more autonomous a system a company wants, the more it must invest in infrastructure that limits that autonomy. Without it, savings in human labor can create disproportionate operational and financial risk. KEY TERMS Runtime control: A technical restriction applied directly while the system is carrying out individual actions. Agent digital identity: A separate identity that allows an autonomous system’s permissions to be precisely controlled and recorded. Agent kill switch: A mechanism that makes it possible to stop an autonomous system immediately when suspicious or dangerous behavior occurs.

Gartner: Corporate policies will not stop an agent. By 2029, 70 percent of organizations are expected to face a serious incident

The full article is available to users registered at Hard Skills