Recent findings from Kore.ai highlight challenges organisations face in managing failures in multi-agent AI environments. While many firms can identify when an AI agent fails, determining the specific agent responsible within a multi-agent setup is often difficult.
A survey of over 400 IT business leaders from companies with 2,000 or more employees found that 70% reported being able to detect a failure. However, identifying which agent caused the issue in a multi-agent environment remains difficult.
Detection Time
The time required to identify agent-related issues varies. Around half of organisations reported detecting problems within one to four hours. A further one third (33%) said it takes between four and eight hours to identify a failure.
Detection Measures
Organisations use a range of methods to detect issues. The most commonly reported approaches include real-time dashboards (39%), automated threshold alerts (29%), and log analysis with pattern detection (17%). In addition, 15% of respondents said they rely primarily on end-users to report issues, meaning some problems are first identified externally.
Trust Levels
The findings also indicate that 53% of firms operate AI agents that they do not fully trust or understand, which can add complexity to both detection and resolution processes.
Cost of Correction
Once an issue is identified and analysed, the remediation process is often manual and resource-intensive. 79% of respondents indicated that correcting autonomous actions requires manual intervention. Of these, 93% reported that the process is costly and disruptive to operations.
Overall, the findings suggest that enterprises may be accumulating “governance debt”, where delays in identifying and addressing issues can increase the cost and impact of incidents over time.
