Enterprise AI incident costs topped $2 million for 43% in WitnessAI survey

WitnessAI’s 2026 Enterprise AI Risk Survey, released July 22, found that 43% of respondents reported enterprise AI incident costs of $2 million or more from all AI-related security incidents over the previous year. The vendor-sponsored survey also put IT and infrastructure at the top of reported shadow-AI activity, a notable result for teams usually expected to enforce technology controls.

The results come from 300 decision-makers at organizations with at least 1,000 employees: 200 held VP, SVP or EVP roles, and 100 were C-suite executives. They are self-reported findings from a security vendor’s survey, rather than a measured average cost across enterprises or an independently verified estimate of the market.

The $2 million figure is an annual aggregate, not the cost of every AI incident. WitnessAI separately said 21% of respondents put the cost of their single most significant AI-related security incident at $1 million or more. Another 17% estimated total annual costs from AI-related incidents at between $10 million and $24.9 million.

WitnessAI said 86% had investigated at least one AI-related security or operational incident in the past 12 months. Ninety-one percent reported concern that AI agents were increasing their exposure to financial risk.

What is shadow AI in an enterprise?

Shadow AI is the unsanctioned use of AI tools by people working at a corporation. In WitnessAI’s survey, 47% named IT and infrastructure as the largest single source of that activity, ahead of sales, business development and marketing. The survey does not establish why IT and infrastructure ranked first, or measure the underlying volume of unauthorized use by department.

The finding arrives as more companies put autonomous-action agents into use. Seventy percent of respondents said their organizations were already using or piloting such agents, while only 18% said every agent had been formally inventoried and approved by the security team.

Where are enterprise AI controls falling short?

Monitoring and ownership were recurring gaps in the responses. Among organizations with deployed agents, 49% reported continuous monitoring, while 12% said they had minimal or no oversight. The survey measures reported controls and perceptions; it does not determine whether the remaining organizations’ approaches are effective.

Senior executives and deployment leaders also described different confidence levels. Sixty-eight percent of C-suite respondents said they had full confidence in their visibility into AI tools and agents, compared with 46% of VPs executing deployments. That 22-percentage-point gap reflects confidence reported by the two groups, not a test of their actual visibility or an explanation for the difference.

Accountability was similarly unsettled. Thirty percent named the CIO or IT leader as primarily responsible for handling AI risk, versus 15% who named the CISO or information-security organization. When asked who would be primarily liable for financial or regulatory harm caused by an AI agent, 26% selected the CIO and 6% selected the CISO.

Fifty-four percent said their organizations devote 21% to 45% of AI budgets to risk management and governance. That spending did not resolve the ownership question in the survey, but the inventory, approval and continuous-monitoring figures identify the specific operational controls respondents reported as incomplete.

These results should not be compared directly with IBM’s $4.99 million global average data-breach figure, which covers a different category of incident. WitnessAI’s number covers respondents’ total costs from AI-related security incidents over a year.

This story draws on original reporting from CIO Dive.