As enterprises move beyond experimental AI pilots and into production with agentic systems, they are colliding with a set of formidable barriers that no keynote presentation fully captures. At VentureBeat’s AI Impact event, Brian Gracely, senior director of portfolio strategy at Red Hat, delivered a stark assessment of what happens when AI agents leave the lab and enter the real world of corporate budgets, security constraints, and human resistance. The challenges span cost discipline, vulnerability management, and the cultural friction that can either accelerate or stall adoption.
Enterprise AI Agents Reveal Cost, Security, and Culture Barriers as Scale Becomes the Real Test
Many enterprise leaders fear they are falling behind in the race to deploy AI agents at scale. Gracely argued that this anxiety is largely misplaced. Organizations typically move up the learning curve faster than they anticipate once they start building. The real problem emerges not from being late, but from being unprepared for the consequences of early success.
As agent usage expands, AI costs rise with equal speed. The cost of inference, token consumption, and GPU compute shifts from an engineering concern to a recurring boardroom discussion. The scale of agentic AI usage is orders of magnitude higher than was typical during the earlier chatbot era. This growth, combined with a growing awareness of dependence on a small number of model providers, is driving enterprises to explore alternatives that offer greater control over both cost and infrastructure.
Gracely pointed directly at the economics of the major model providers. “The two or three top providers are already telling the market that they’re losing money, and they’re trying to go public to make up those gaps,” he said. “At some point, the dependency on that means you’re either going to buy at a very high-cost level, or you’re going to figure out alternatives to control what you’re doing.”
Right-Sizing Models Is the Fastest Lever for Cutting Agent Costs
The most common and most expensive mistake enterprises make is defaulting to the most capable model available for every task, regardless of its actual complexity. This overspending is avoidable. “If I’m simply trying to resolve an insurance claim, I don’t need to know about the history of Western civilization in my model, I don’t need to know World Cup soccer scores,” Gracely said.
The technical mechanism many organizations are adopting to solve this is semantic routing. This technique automatically classifies incoming requests and routes each one to a model appropriately sized for the task, removing the need for users to manually select a model. Caching repetitive queries further reduces the number of requests that need to reach GPU compute. Together, these tools challenge the assumption that efficiency and innovation must be trade-offs.
“There’s a lot you can do at a GPU infrastructure level, and quite a bit you can do in terms of flexibility of models,” Gracely explained. “Those give excellent choices in terms of the levers you’re trying to pull, whether you need efficiency or you need innovation. That shouldn’t be a binary choice.”
The financial discipline required for token spend mirrors the FinOps practices that took years to mature in the cloud computing era. Those underlying frameworks will transfer even as the vocabulary changes, particularly as organizations push for internal education on model selection so teams stop defaulting to the most prominent option for tasks that do not need it. “The same way we first had to teach the financial people what an EC2 instance is and what an S3 bucket is, you’re going to have to start explaining tokens to them,” Gracely said. “We don’t always need a Rolls-Royce. We don’t always need caviar, because we’re trying to do basic types of things.”
AI-Driven Vulnerability Discovery Is Compressing Patch Windows
AI-powered vulnerability discovery is forcing a fundamental rethinking of patch management cycles. Attackers can now use AI to uncover and exploit security flaws much faster than in the past, meaning traditional patch schedules may no longer be sufficient.
“Most companies are probably going to have a window of somewhere between seven and 14 days to stay ahead,” Gracely said. “There are groups, Red Hat included, that are going to build patches for these, but the embargo window is going to be short.”
The nature of what defenders need to monitor is also changing. Rather than simply hunting for isolated critical flaws, AI security tools can now identify combinations of seemingly minor vulnerabilities that become dangerous only when chained together. As both software complexity and vulnerability discovery accelerate, the ability to rapidly manage and update software is transforming from an operational task into a strategic capability.
Subject Matter Expert Buy-In Determines Whether Agents Scale
The single most overlooked barrier to scaling AI agents is the human factor. Organizational adoption hinges on deep, sustained involvement from the subject matter experts whose knowledge the agent is meant to encode. Earning their buy-in is a prerequisite, not an afterthought.
“You have to think about the incentives, what you do for people who participate in this work so they don’t feel threatened that it’s going to take away their job, and how you incentivize people in the long run to cooperate with that innovation,” Gracely said.
The risk for enterprises is that even the most technically sound agent will fail if the people whose expertise it relies on are not motivated to participate. Successful adoption requires aligning incentives so that collaboration is seen as an opportunity to shape the future of work, not a threat to existing roles.
What Enterprises Should Do Now
The path to production-scale agentic AI requires more than technical sophistication. Enterprises must build cost discipline through model right-sizing and semantic routing, accelerate patch management cycles to keep pace with AI-driven vulnerability discovery, and invest heavily in the cultural and incentive structures that bring subject matter experts into the process as collaborators rather than skeptics. The organizations that address all three dimensions simultaneously will be the ones that actually scale.
