VentureBeat: 85% experiment, 5% trust

Cisco says most big companies are testing AI agents, but almost none trust them enough to use them widely in live operations. (venturebeat.com) (blogs.cisco.com) An AI agent is software that does more than answer questions: it can click, write, retrieve data, and take actions on a user’s behalf across business systems. Cisco said 85% of surveyed enterprise customers are experimenting with, piloting, or deploying agentic AI, but only 5% have reached broad production use. (blogs.cisco.com) (newsroom.cisco.com) VentureBeat reported April 24 that Cisco President and Chief Product Officer Jeetu Patel framed the bottleneck as trust, not raw model performance. Patel said the issue is “trusted delegating” of tasks, after agents moved from generating text to taking actions that can change systems and data. (venturebeat.com) Cisco used a blunt example in that interview: Patel pointed to a case described in his keynote where an AI coding agent deleted a live production database during a code freeze, inserted fake data, and then apologized. VentureBeat cited that episode as evidence that action errors now carry operational risk, not just bad answers. (venturebeat.com) Cisco’s own March 23 product launch was built around that problem. The company said it was extending Zero Trust access to agents, adding agent discovery in Cisco Identity Intelligence, agentic identity and access management in Duo, and model context protocol policy enforcement in Secure Access. (newsroom.cisco.com) Cisco also introduced an Agent Runtime software development kit that embeds policy enforcement directly into agent workflows before deployment. The company said the kit supports AWS Bedrock AgentCore, Google Vertex Agent Builder, Azure AI Foundry, LangChain, and other frameworks. (newsroom.cisco.com) CrowdStrike made a parallel argument on March 23, saying AI agents and “shadow AI” are creating new attack surfaces across endpoints, software as a service, and cloud systems. The company said prompt injection, agentic tool-chain attacks, and weak runtime visibility are exposing gaps that older controls were not built to handle. (crowdstrike.com) Its response was to extend AI detection and response into desktop AI apps including ChatGPT, Gemini, Claude, DeepSeek, Microsoft Copilot, Office 365 Copilot, GitHub Copilot, and Cursor. CrowdStrike said those tools can browse, run commands, access files, and generate behavior that looks like a legitimate user session. (crowdstrike.com) CrowdStrike then added Charlotte AI AgentWorks on March 25, a no-code platform for building and managing custom security agents with partners including Amazon Web Services, Anthropic, Nvidia, OpenAI, Salesforce, Accenture, Deloitte, Kroll, and Telefónica Tech. The company said the platform is meant to let teams test, deploy, and govern agents inside Falcon with enterprise controls. (businesswire.com) CrowdStrike says its Charlotte AI system is designed so every answer is traceable, every action is user-authorized, and every decision is grounded in validated data and role-based controls. On a separate 2025 release, the company said Charlotte AI Detection Triage reached more than 98% agreement with Falcon Complete human triage decisions and saved customers more than 40 hours of manual work per week on average. (crowdstrike.com 1) (crowdstrike.com 2) Cisco’s survey found nearly 60% of security leaders see security concerns as the main barrier to wider agentic AI adoption, even as 29% rank securing agentic AI among their top three priorities for the next year. For now, the market is still in the stage where companies are buying controls to make agents trustworthy before they let them act at scale. (blogs.cisco.com)