Kore.ai has warned that companies are adding AI agents faster than they can manage them, after new research found that 72% of enterprises say their agents operate with unmanaged financial or compliance risk.
The 2026 Kore.ai Agent Productivity Index surveyed more than 400 IT business leaders on the state of agentic AI in the enterprise. It found that companies are giving AI agents authority over data, decisions, and customer interactions while many leaders remain unable to fully trace or trust how those agents use that authority.
Among respondents, 79% said they had had to reverse an action taken by an AI agent, 70% had faced a failure their teams could not trace, and 62% had delayed deployments because of governance concerns. More than half, 53%, said they were running agents they did not fully trust or understand, while 42% reported lost revenue tied to an AI agent failure.
The survey also found that 40% of enterprises had seen a single agent failure cascade across multiple systems. In the first half of 2026, 41% of agents were running data migrations and system updates, 26% were approving or denying decisions, and 15% were acting on financial transactions.
Raj Koneru, CEO and founder of Kore.ai, said: “Enterprise AI has shifted from showing that AI works to proving it can be trusted. Governance has to be built into the agent itself, not added once it is running, because trust comes from visibility, reproducibility, auditability, and control, not from the model getting it right every time.”
He added: “The companies that scale AI will be the ones using AI to build, govern, and improve AI on a single layer. That is the architecture this market is moving to, and it is the one we’ve built.”
The report argues that adding guardrails, monitors, or policy engines after an agent has been built is structurally weaker than designing governance into the agent lifecycle. Kore.ai said post-build controls govern the running agent, but cannot shape how it was built or how it will be updated.
The company describes its Kore.ai Agent Platform, Artemis edition, as a single system for building, deploying, managing, and improving agents. Its AI agent architect, Arch, turns plain-language objectives into agents defined in Agent Blueprint Language, with governance, observability, and operational control enforced before deployment.
Peter Mullen, chief marketing officer at Kore.ai, said: “The market is solving for visibility, but enterprises need accountability. Those are not the same thing. An agent that can be watched but not governed is still a liability.”
Agentic AI is now entering operational environments. Unlike earlier generative AI tools that mainly produced text, code, summaries, or analysis for human review, agents can take action across systems. A flawed answer is one category of risk; an autonomous action affecting customers, data, finance, or operations is another.
The same governance pressure has been visible at board level. In directors put AI governance on agenda, polling from the Institute of Directors showed companies beginning to formalise oversight of AI. Kore.ai’s findings show how quickly the control problem intensifies once agents gain operational authority.
The enterprise challenge is partly architectural. Agents need access to data, APIs, workflow tools, customer systems, finance platforms, service desks, and decision rules. Each connection increases usefulness but also raises the cost of failure. If an agent cannot be traced, tested, constrained, or rolled back, productivity gains can quickly create operational exposure.
Governance therefore needs to cover design, deployment, monitoring, escalation, and continuous improvement. Companies need to know what an agent is allowed to do, which data it can access, which human approvals are required, how decisions are logged, how exceptions are handled, and who is accountable when something goes wrong.
The revenue loss figure in the survey is particularly notable because it shifts the debate beyond theoretical risk. Enterprises are already experiencing commercial consequences tied to agent failures. Some organisations may slow adoption, while others will invest more heavily in control layers, testing, and auditability.
AI agents remain attractive because the productivity prize is substantial. They can compress workflows, reduce manual handoffs, improve service speed, and operate across systems in ways that conventional automation cannot. Those advantages depend on whether governance is designed before authority is granted.
Kore.ai’s research points to a market now concerned with AI operations rather than experimentation alone. The next test is whether enterprise leaders can scale agents without losing visibility over the decisions and actions those agents perform.
You must be logged in to post a comment.