Enterprise AI agents are moving from pilot programs to production, but the transition is uncovering significant gaps in cost management, security patching, and organizational culture. According to Brian Gracely, senior director of portfolio strategy at Red Hat, companies often overestimate how far behind they are in deployment while underestimating the operational friction and financial volatility that occur once these autonomous systems scale.
The shift toward agentic AI—systems capable of planning and executing multi-step tasks independently—requires a level of token consumption and compute power orders of magnitude higher than the chatbot era. This surge in usage is transforming AI spend from a technical line item into a recurring boardroom discussion, as organizations grapple with a heavy reliance on a small number of model providers.
While many leaders fear they are losing a competitive race, Gracely notes that teams typically move up the learning curve faster than expected. However, this rapid progress often leads to “over-provisioning,” where companies default to the most capable model available for simple tasks that do not require such vast computational resources.
Solving the AI Cost Crisis Through Model Right-Sizing
The primary driver of escalating AI costs in the enterprise is the tendency to use the most powerful model available regardless of the task’s complexity. Gracely explains that using a massive model to resolve a routine insurance claim is inefficient, as that specific task does not require the model to possess broad general knowledge, such as the history of Western civilization or sports scores.
To combat this, enterprises are implementing semantic routing. This mechanism automatically classifies incoming requests and directs them to a model sized specifically for the complexity of the task, removing the need for users to manually select a model. Additionally, infrastructure techniques like caching repetitive queries are being used to reduce the frequency with which requests must reach GPU compute, lowering the overall cost per interaction.
This approach mirrors the evolution of FinOps practices in cloud computing. Just as organizations spent years refining how they managed EC2 instances and S3 buckets to control cloud spend, they must now educate financial teams on tokens. The goal is to move away from a “Rolls-Royce” approach to every task, prioritizing efficiency over raw power when the objective is basic operational execution.
The Shrinking Window for AI-Driven Vulnerability Patching
The rise of AI agents is fundamentally altering the timeline for cybersecurity. AI-powered tools are now capable of discovering software vulnerabilities faster. This acceleration means that long-established patch management cycles may no longer be fast enough.
According to Gracely, the window for companies to identify, validate, and deploy patches to stay ahead of attackers is likely between seven and 14 days. While groups including Red Hat build the necessary patches, the “embargo window” is going to be short.
Security teams are also facing a shift in what they need to look for. Rather than searching for a single, critical flaw, AI security tools can now identify combinations of seemingly minor vulnerabilities that become dangerous only when chained together. Consequently, the ability to rapidly manage and update software is becoming a strategic capability rather than simply an operational one.
Overcoming Organizational Friction and Expert Resistance
The technical ability to deploy an AI agent does not guarantee its success; the final hurdle is often the human element. For an AI agent to scale, it must encode the deep, sustained involvement from the subject matter experts whose knowledge the agent is meant to encode. If these experts view the agent as a threat to their job security, they are unlikely to provide the high-quality data and guidance necessary for the system to function.
Gracely argues that earning the buy-in of subject matter experts and compliance teams is a prerequisite for scaling, not an afterthought. This requires a shift in corporate incentives. Organizations must find ways to reward the experts who participate in this work, ensuring they feel like partners in innovation rather than targets for replacement.
Without this cultural alignment, agents remain trapped in “pilot mode,” unable to move beyond early champions because the people who actually understand the business logic refuse to cooperate with the development process.