72% of enterprises say their
AI agents operate with unmanaged risk.
AI agents are everywhere.
Governance is not.
Eight in ten enterprises have had an AI agent execute a consequential action in production and faced real costs correcting it. We set out to understand why by examining how organizations detect failures, establish accountability, and make governance decisions about AI agents they don't fully understand.
Deploying AI agents is easy. Governing them in production at scale is the biggest challenge.
Organizations have governance frameworks and policies defined, yet continue to experience costly reversals, slow failure attribution, and deployments that proceed without a complete understanding of agent behavior. Governing agents in production is where those frameworks are falling short. This research measures exactly how wide that gap is.
Teams can see that something is broken.
Finding out what is broken takes far too long.
Enterprises have invested in detection. What most have not built is traceability. In multi-agent environments, the gap between flagging a failure and attributing it to a specific agent has a direct business cost: every hour that attribution takes is an hour the problem continues to run impacting your business outcomes.
Most organizations know when something has gone wrong, but seven in ten cannot identify which agent in a multi-agent environment caused it. Observability and attribution are different capabilities, and the data shows enterprises have invested in one without building the other.
Agent failures do not stay in the IT layer. They reach revenue, customers, and contracts.
When agents fail in production, the consequences are not contained to engineering. Most of that exposure is still being classified as an IT incident, which means the true cost never surfaces where decisions get made.
Agent failures increasingly affect revenue, contractual obligations, and customer relationships. Yet many organizations continue to manage them primarily within IT, limiting visibility into their broader business impact.
Seeing this in your own production environment?
Kore.ai Artemis traces every agent action, enforces policy at runtime, and contains failures before they reach revenue. See how the harness governs the full agent lifecycle.
of consequential autonomous actions required manual reversal. Of those, 93% were described as costly and disruptive, meaning the act of correction compounded the original failure.
Reversal rates remained high despite safeguards being in place.
Organizations have invested heavily in governance safeguards, yet the data suggests those safeguards do not always translate into effective outcomes. As agents take on more consequential responsibilities, the ability of governance mechanisms to prevent, contain, and recover from failures becomes increasingly important.
of enterprises report agents have autonomously executed consequential actions in production.
Financial transactions, workflow approvals, and data migrations are increasingly being executed by agents, often with limited human oversight.
As agents take on more consequential responsibilities, governance effectiveness depends less on the presence of controls and more on their ability to prevent costly outcomes.
Governance frameworks are defined, yet leadership confidence remains low.
Organizations report widespread adoption of governance mechanisms, including kill switches, containment procedures, and executive review processes. Yet confidence remains low, with 72% of respondents believing autonomous agents introduce unmanaged financial or compliance risk.
As agent ecosystems become more interconnected, governance effectiveness increasingly depends on containment. The ability to pause an agent is valuable, but limiting the spread of failures across connected systems remains the greater challenge.
About this research
This report is based on a survey of 408 IT and engineering leaders in enterprise organizations. Every respondent is actively running AI agents in production.
Questions covered six governance dimensions: failure detection, autonomous action accountability, governance and containment infrastructure, executive oversight, standardization, and model cost optimization.
How Artemis moves governance from oversight to execution
The survey findings point to a common challenge: organizations that have invested in AI agents are now finding that deployment was the easy part. Accountability gaps, safeguards that cannot be verified, and costs that scale with usage instead of value are not engineering problems. They are governance problems. And governance cannot be retrofitted.
Unlike traditional AI platforms, Artemis treats governance as part of the agent lifecycle, not a layer added on top of it. Every capability described below is built into how agents are defined, validated, deployed, and managed at scale.
Embed governance within architecture
Artemis enforces governance at the structural level, not the prompt level. Policies, constraints, and escalation rules are compiled into the agent definition before it ever reaches production. There is no configuration to drift and no prompt to override.
Enforce policy at runtime
A deterministic runtime engine enforces boundaries on every agent action in real time. If an agent attempts something outside its defined scope, the system intercepts it before it executes. Safeguards are not monitored after the fact. They hold in the moment.
Trace every decision
Every agent action, handoff, and constraint check is logged end to end across 100% of interactions. When something goes wrong, the answer is in the trace. Not reconstructed from assumptions, not inferred from outputs. The decision chain is on record.
Govern the system, not the model
Most governance approaches treat the LLM as the unit of control. Artemis governs the system: the agent, the tools it can call, the data it can access, and the actions it is permitted to take. The model is one component. The system is what enterprises are accountable for.
Manage the lifecycle as one system
Artemis covers the full agent lifecycle: design, testing, deployment, monitoring, and updates in a single governed environment. There is no handoff between tools where policies get lost or documentation falls out of sync with what is actually running.
Scale without multiplying oversight
Governance in Artemis is inherited, not applied manually. When an organization scales from 10 agents to 1,000, the same policies apply without requiring proportionally more review or configuration. Governance scales with the deployment, not against it.
“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.”
for the agentic enterprise
Take the report with you
Get the complete Kore.ai Agent Productivity Index as a PDF, ready to share with your leadership team, your board, or anyone deploying agents in production.
This report presents findings from a proprietary survey of enterprise IT and engineering leaders, commissioned by Kore.ai and conducted by Propeller Insights. Figures are self-reported and rounded for presentation; percentages reflect the share of respondents selecting each option, and multi-select questions may total more than 100%. The findings are provided for general informational purposes only and do not constitute legal, financial, or professional advice. Product capabilities referenced are subject to change. Kore.ai Artemis, Arch, and Agent Blueprint Language are trademarks of Kore.ai. © 2026 Kore.ai. All rights reserved.
