US Firms Rush AI Agents, Few Report Returns, Many Breaches

A survey of 1,200 top executives and 1,200 employees found 97% of U.S. companies deployed AI agents during the past 12 months and 52% of employees are using them. Fewer than three in 10 organizations report meaningful financial returns from generative AI, and 59% spend more than $1 million a year on AI technology. Fifty-four percent of top executives said the effort is straining their organizations.

Security incidents tied to unapproved AI use were common. Two-thirds of executives reported data leaks or security breaches after employees used unapproved AI tools. Thirty-five percent of employees acknowledged entering company secrets into public AI services. The survey found 36% of companies do not have a formal plan to oversee AI agents, and 35% said they could not immediately shut down an agent that behaved unexpectedly.

Companies reported problems with strategy and adoption. Seventy-nine percent said they face challenges adopting AI, a rise from the prior year. Seventy-five percent of executives described their AI strategy as existing more for public image than for directing employees, and 48% called their AI efforts a “massive disappointment.” Thirty-nine percent reported no formal plan to generate revenue from AI tools.

A separate analysis of 200,000 user interactions, 3,000 demo requests and more than 2,000 conversations with business and technology leaders found 62% of companies lack a clear starting point for AI agent work, 41% treat AI projects as side initiatives, and 32% stall after pilot programs without scaling to production. Only 23% of respondents reported substantial returns specifically from AI agents. Usage data showed 70% of employees and 94% of top leaders use AI tools at least 30 minutes per day, and 64% of executives spend two or more hours daily with AI tools.

Participants at recent technology events in California highlighted technical and cost challenges. Kevin McGrath, founder of startup Meibel, warned that companies route too many tasks through costly systems and quoted an example of an “AI Claw bot” wasting millions of tokens. Deep Shah, a Google engineer, pointed to high operating costs and described poorly designed implementations as systems that burn cash rather than reduce expenses. He added that unclear responsibility for agent oversight creates wider organizational risk.

Chris Han, a leader at ThinkingAI, said popular consumer-grade tools cannot meet complex corporate needs. He noted companies must address engineering issues such as memory management, coordinating multiple agents and agent-to-agent communication before deployments can deliver consistent productivity gains.

In response to reported breaches, some firms are considering monitoring and shutdown protocols for AI agents, according to the survey. Over the past year, enterprises increased experimentation with generative models and autonomous agents, launching pilot programs across functions. Many projects stalled at pilot stages or were launched without clear use cases, governance frameworks or cost controls, the research found.