The hidden challenges slowing enterprise AI agents

Enterprises frequently misjudge their lag in agentic AI adoption

Many executives believe they are already behind in AI because every conference, keynote, and product announcement creates the impression that competitors have already solved enterprise AI. That perception is usually inaccurate. According to Brian Gracely, Senior Director of Portfolio Strategy at Red Hat, companies often move up the AI learning curve much faster than they expect once they start building real applications.

The more important question is not whether you started first. It is whether you can operate AI at scale. That is where the real work begins. Production systems introduce problems that never appear during pilots. AI costs grow rapidly as usage expands. Security risks become more complex because autonomous agents make decisions and interact with business systems without constant human involvement. At the same time, organizations discover that technology is only one part of the challenge. Governance, business processes, and employee engagement become equally important.

This is why many successful AI projects slow down after early demonstrations. The technology may work exactly as expected, but the organization has not yet adapted to support it. Teams need clear ownership, financial oversight, security policies, and executive sponsorship. Without those elements, AI remains a collection of isolated projects instead of becoming part of the business.

Another issue is dependence on a small number of large AI model providers. Gracely pointed out that leading providers are already telling the market they are operating at a loss while searching for sustainable business models. For enterprise leaders, this raises strategic questions about future pricing, long-term vendor dependence, and infrastructure flexibility. Many organizations are therefore evaluating open-source models, hybrid deployments, and multi-model strategies that provide greater control over costs and reduce reliance on a single provider.

The companies that will create lasting value from AI are unlikely to be the ones making the loudest announcements today. They will be the ones building operational discipline early. AI is becoming another core business capability. Like cloud computing before it, success depends less on initial adoption and more on the ability to manage it consistently over time.

Right-sizing AI models is critical for managing escalating operational costs

One of the fastest ways to reduce AI costs is surprisingly simple: stop using the largest model for every task.

Many organizations automatically select the most capable model because it appears to be the safest option. In reality, most enterprise workloads do not require the highest-performing model available. Brian Gracely explained this clearly when discussing routine business processes such as insurance claims. These tasks require domain-specific reasoning, not broad knowledge across countless unrelated subjects. Using an oversized model for straightforward work increases costs without creating additional business value.

Modern AI systems allow companies to make smarter decisions automatically. Semantic routing classifies each request and sends it to the model that best matches its complexity. Straightforward requests can be handled by smaller, less expensive models, while complex reasoning tasks are directed to larger models only when necessary. The user experiences a seamless service, while the organization significantly reduces token consumption and computing costs.

Infrastructure optimization provides another layer of savings. Many enterprise AI requests are repetitive. Instead of generating a new response every time, organizations can use caching to store previous results and reuse them when appropriate. This reduces the number of expensive GPU computations required, lowering operating costs while improving response speed.

Gracely emphasized that efficiency and innovation should not be viewed as competing priorities. Organizations have multiple technical and operational levers available, including model selection, routing strategies, and infrastructure optimization. The goal is not simply to reduce spending. It is to allocate AI resources where they create the greatest business impact.

This also changes how finance and technology teams work together. AI spending is increasingly measured through token usage and model consumption rather than traditional infrastructure metrics alone. Gracely compared this shift to the early days of cloud computing, when finance departments had to learn concepts such as Amazon EC2 instances and Amazon S3 storage before effective cloud cost management became possible. The same evolution is now taking place with AI.

For executives, this means AI governance should include financial governance from the beginning. Teams need clear policies on model selection, usage monitoring, and cost accountability. Without these controls, AI expenses can grow much faster than expected as adoption expands across business units. With them, organizations can scale AI confidently while maintaining predictable economics and preserving flexibility for future innovation.

Establishing AI financial management practices similar to FinOps is essential

As AI moves from experimentation to enterprise-wide deployment, financial management becomes a strategic capability. Many organizations begin by measuring technical performance, but long-term success depends just as much on understanding how AI consumes resources and how those resources translate into business value.

Brian Gracely, Senior Director of Portfolio Strategy at Red Hat, argued that enterprises are entering a stage that closely resembles the early years of cloud adoption. At that time, finance teams had to learn new concepts before they could effectively govern cloud spending. AI introduces a similar challenge. Executives, finance leaders, and business managers now need to understand metrics such as token consumption, model usage, and inference costs so they can make informed investment decisions.

This is more than introducing new terminology. Every interaction with an AI model has a cost, and those costs accumulate rapidly as AI agents perform more work across multiple business functions. Without visibility into usage patterns, organizations can easily discover that successful adoption is accompanied by unexpectedly high operating expenses.

Companies should therefore establish governance frameworks that treat AI spending as an ongoing operational expense rather than an isolated technology budget. This includes defining who approves model selection, monitoring consumption across departments, setting spending thresholds, and regularly reviewing whether workloads are running on the most appropriate models. Financial oversight should become part of everyday AI operations instead of being added after costs begin to rise.

Education also plays an important role. Engineers understand technical performance, while finance teams focus on budgets and returns. Those perspectives need to come together. Gracely emphasized that organizations will need to explain AI concepts, including tokens, to financial stakeholders so spending decisions are based on a shared understanding of how AI systems operate.

The companies that manage AI costs effectively will not necessarily spend less. They will spend with greater precision. They will understand where AI creates measurable business value, where lower-cost models are sufficient, and where additional investment produces meaningful returns. That level of discipline allows organizations to scale AI with confidence rather than constantly reacting to unexpected expenses.

Accelerated vulnerability discovery through AI demands rapid patching and agile security operations

AI is changing cybersecurity at a much faster pace than many organizations expected. It is making vulnerability discovery faster, increasing the speed at which weaknesses become known, and reducing the amount of time defenders have to respond.

Brian Gracely warned that traditional patch management cycles may no longer be fast enough. Organizations that previously measured patch deployment over several weeks may now face much tighter timelines. He estimated that many companies will have only about 7 to 14 days to identify, validate, and deploy patches before attackers can take advantage of newly discovered vulnerabilities.

This shift affects executive priorities as much as technical operations. Security teams can no longer depend solely on periodic maintenance schedules. Continuous monitoring, rapid validation, and automated deployment processes are becoming essential parts of enterprise resilience. Organizations that reduce response times strengthen their security posture and business continuity.

AI is also changing the nature of cyber risk itself. Instead of identifying only a single critical vulnerability, AI-powered security tools can uncover combinations of smaller weaknesses that become dangerous when exploited together. These chained vulnerabilities are more difficult to detect through traditional manual analysis because each issue may appear relatively minor on its own. AI can identify these relationships much faster, allowing both defenders and attackers to move more quickly.

This creates a new expectation for software governance. Security is no longer only about preventing incidents. It is about maintaining systems that can be updated continuously without disrupting business operations. Organizations with modern software delivery processes will generally be better positioned to respond as the pace of vulnerability discovery continues to accelerate.

For executive leadership, cybersecurity should increasingly be viewed as an operational capability that supports growth rather than simply a compliance requirement. AI adoption expands business opportunities, but it also expands the attack surface. Companies that invest in faster patch management, stronger governance, and automated security processes will be better prepared to capture AI’s benefits while managing its risks.

The scalability and success of enterprise AI agents depend on organizational buy-in and expert engagement

Technology is rarely the biggest obstacle to scaling AI. People are.

Many organizations can build an effective AI agent, but far fewer can integrate it into daily business operations across multiple teams. That difference usually comes down to whether the people with the deepest operational knowledge actively support the initiative. Brian Gracely, Senior Director of Portfolio Strategy at Red Hat, emphasized that subject matter experts are not optional contributors. They are essential because their expertise determines how accurately an AI agent performs in real business situations.

These experts understand the processes, exceptions, regulatory requirements, and customer expectations that are often undocumented. If they are involved only at the beginning of a project, or not at all, the resulting AI system may perform well during testing but struggle once deployed in production. Continuous collaboration helps keep AI systems aligned with changing business requirements and improves their reliability over time.

Organizations also need to recognize that AI can create uncertainty among employees. If people believe their expertise is simply being extracted and automated, participation will naturally decline. Gracely argued that leaders should think carefully about incentives and ensure employees understand how AI supports their work rather than replacing their contribution. Building trust is a leadership responsibility.

Successful companies typically establish clear governance around AI ownership and accountability. Business leaders, technical teams, compliance functions, and subject matter experts all have distinct responsibilities throughout the AI lifecycle. This shared ownership improves decision-making and helps organizations identify operational, legal, and regulatory issues before they become larger problems.

Compliance teams also play a central role as AI adoption expands. Enterprise AI agents increasingly interact with sensitive business information, customer data, and regulated processes. Early involvement from compliance, legal, and risk management teams allows organizations to design appropriate controls from the beginning instead of introducing them after deployment. This reduces implementation delays and strengthens confidence in AI-driven decisions.

For executives, AI transformation should be viewed as both a technology initiative and an organizational change initiative. Investment in models and infrastructure is necessary, but it is not sufficient. Long-term success depends on creating an environment where business experts, technical teams, finance, security, legal, and leadership work toward shared objectives. Companies that achieve that alignment will be in a stronger position to scale AI beyond isolated pilots and embed it into core business operations.

As Gracely noted, organizations should consider “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.” That perspective highlights an important reality: sustainable AI adoption is built through engagement, trust, and collaboration across the enterprise.

Main highlights

• Focus on operational readiness: Many companies overestimate how far behind they are. Leaders should prioritize governance, cost management, security, and organizational alignment because these factors determine whether AI scales beyond successful pilots.
• Match the model to the business task: Using the largest AI model for every workload increases costs without improving outcomes. Adopt semantic routing, smaller models for routine tasks, and infrastructure optimizations such as caching to improve efficiency while maintaining performance.
• Build financial discipline into AI from day one: AI spending should be managed with governance similar to cloud FinOps. Educate finance and business leaders on token usage, monitor AI consumption, and establish clear policies for model selection to keep costs predictable as adoption grows.
• Modernize security for AI-speed threats: AI is accelerating vulnerability discovery, shrinking patch windows to as little as 7 to 14 days, according to Brian Gracely of Red Hat. Invest in faster patch management, continuous monitoring, and automated security processes to reduce business risk.
• Make people central to AI strategy: Enterprise AI succeeds when subject matter experts, compliance teams, and business leaders actively shape deployment. Create incentives, build trust, and establish clear ownership so AI becomes a sustainable business capability rather than a series of isolated projects.