The ease of creating AI agents through platforms like Anthropic's Claude Cowork is leading to an unmanaged proliferation within enterprises, presenting significant cybersecurity, management, and cost challenges. An overabundance of AI agents, particularly when multiple agents perform redundant tasks, is creating complex problems for corporate IT departments.
As AI agents become widely adopted in the corporate world, the issue of their uncontrolled spread is becoming more pronounced, leading to management chaos, security vulnerabilities, and rising operational costs. Major companies are now scrambling to address these challenges. While AI agents are being rapidly adopted, some forward-thinking companies are already taking steps to mitigate the risks associated with having too many of them. Companies including Lyft, DaVita, and GitLab report that they are currently managing the challenges posed by a surge in AI agents. Their goal is to control this unchecked expansion without dampening employee enthusiasm for using AI technology. This phenomenon, termed AI agent sprawl, is partly driven by platforms like Anthropic's Claude Cowork, which allow employees without technical backgrounds to easily create independent AI assistants. The proliferation is further accelerated by open-source tools like OpenClaw, which facilitate the orchestration of multiple agents. However, businesses warn that an excessive number of agents, especially those performing duplicate functions, introduces cybersecurity risks and management headaches for IT teams, while also significantly driving up computational costs. Michael Friedlander, CIO for the Americas at ice cream maker Magnum, commented, "When anyone can create an AI agent with a few clicks, you inevitably end up with many employees having agents that do the same thing." Friedlander stated that the ice cream giant will eventually need to consolidate and centrally manage the vast number of internally built AI agents, while striving not to stifle employee innovation. "This will involve issues like token billing and cost management. Ultimately, we must figure out how to govern this properly to ensure overall spending meets financial compliance requirements." Data from research and advisory firm Gartner indicates that within the next two years, the average Fortune 500 company will be running over 150,000 AI agents. Yet, only 13% of companies believe they have established a robust governance mechanism for these agents.
**Managing the Risks of Sprawl** Currently, workplace AI agents are primarily used for basic tasks such as writing code, summarizing emails, customer service, and data analysis. More sophisticated agents are being deployed for deep research and end-to-end business process automation. Jay Das, co-founder and president of venture firm Sapphire Ventures, noted that these AI agents might run on employee laptops, corporate servers, or other internal systems, making it difficult for IT departments to track and control them comprehensively. Mike Tuckey, Chief Customer Officer and CIO at credit scoring company FICO, revealed that the firm's 3,500 employees create dozens of new AI agents every day. "New agents are being created almost daily, and they are being used at every level of the company." These agents range from personal bots for email management and report writing to large business intelligence agents that maintain datasets for multiple projects. FICO has now established governance rules to prevent multiple AI agents from producing conflicting results for the same problem. Tuckey stated, "We recognize the potential risks here." Madhu Narasimhan, CIO at DaVita, disclosed that employees at the healthcare provider have created over 10,000 AI agents. For cybersecurity reasons, the kidney care company prohibits consumer-grade AI tools from accessing its corporate network. "We serve patients and must scale AI in a secure and controlled manner." DaVita has also developed an internal platform to regulate AI inference calls on-demand, manage token costs, and allocate more resources to high-performing, efficient agents.
**Centralizing Control Over Unmanaged Bots** A growing number of technology executives are taking proactive steps to manage AI agent sprawl and prevent the problem from worsening. Jason Vogrinec, Executive Vice President of AI Transformation at ride-hailing platform Lyft, said the company has granted all employees access to Claude and has established an IT-compliant skill-sharing mechanism. This includes pre-set instruction templates to help Claude agents handle specific work tasks. Lyft is also building a centralized management platform where IT will oversee all AI agents. Vogrinec noted that as a publicly listed company with multiple regulatory obligations, agent sprawl presents a significant compliance challenge. Anthropic stated that it has launched several enterprise management features, supporting role-based access control, spending management, usage analytics, audit logs, and a curated plugin library – essential capabilities for IT administrators. Manu Narayan, CIO at GitLab, said the company's existing governance rules and risk control mechanisms have effectively curbed the excessive and disorderly expansion of AI agents. In the short term, however, some see a silver lining to the agent sprawl. Narayan added, "Some level of AI agent sprawl is inevitable in the short term, and we are tolerant of it because artificial intelligence itself holds tremendous developmental opportunities."
Comments