Enterprise AI Agents: The Cost and Security Pitfalls

The challenges enterprises face when fully deploying AI agents are being clarified once again. Brian Gracely, Senior Director of Portfolio Strategy at Red Hat, explained the current state of affairs at VentureBeat's AI Impact event from three perspectives: cost management, security, and organizational culture.

Many business leaders are overly concerned about falling behind in AI agents, but according to Gracely, teams that start building often climb the learning curve faster than expected. In other words, the anxiety about "significantly lagging behind competitors" is somewhat exaggerated compared to reality. However, the speed of learning itself creates another problem. As agent usage expands, the cost of AI grows rapidly, and cost management—which should have been an engineering issue—becomes a topic in executive meetings.

Behind the cost increases lies the reality that AI agent usage is "orders of magnitude higher" than during the chatbot era. Moreover, the concentration of dependency on a small number of model providers is also a concern for enterprises. Gracely points out that "the top 2-3 major providers have already announced losses and are aiming for public offerings to cover the gap." He further states that "eventually it will come down to either paying consistently very high costs or finding alternative options and maintaining control ourselves," emphasizing the risks of over-reliance on specific providers.

Gracely explains that the most immediately effective approach to cost reduction is to differentiate AI model usage according to task complexity. As illustrated by the statement "if you're just processing insurance claims, you don't need a model trained on all of Western civilization history and soccer scores," using oversized models for given tasks is the primary driver of unnecessary spending. A mechanism that automates this is "semantic routing," which automatically classifies request content and directs each to an appropriately sized model. Additionally, caching frequently repeated queries reduces the frequency of GPU recalculation.

These techniques demonstrate that the tradeoff between efficiency and performance is not necessarily something to be avoided. Gracely states: "There is much that can be done at the GPU infrastructure layer, and considerable opportunities exist in terms of model flexibility. Whether you want to prioritize efficiency or innovation, you have options to accommodate either," suggesting that maintaining cost discipline while running sophisticated agents through proper design and configuration is a realistic option.

AI agent security presents unique risks distinct from traditional chatbots. Because agents access external systems and make autonomous decisions, the "attack surface" is significantly broader. Gracely warned that organizations are insufficiently aware of these security blind spots inherent in autonomous systems. He also noted that organizational and cultural friction, rather than technical issues, is the primary barrier preventing agent adoption from spreading beyond early adopters to the broader organization.

The overall picture presented by these field insights supports the view that the "next barrier" in enterprise AI is not model performance but operations. The three challenges of cost, security, and organization cannot be resolved by technology selection alone and require concurrent transformation in management, systems, and culture. The difference between enterprises that can transition AI agent deployment from the pilot phase to full-scale implementation and those that cannot is precisely reflected in their capacity to respond in this domain.