Controlling AI models is harder than building them
Before companies safely grant autonomy to AI agents deployed in their business processes, they must first establish an infrastructure of governance, accountability and control.
As AI systems do more than generate content, businesses are discovering that controlling AI models after deployment is more challenging than building them.
AI agents can get held up in loops. Errors can move through workflows before someone catches them. And when something goes wrong, businesses struggle to determine exactly what happened and why. As a result, attention is increasingly shifting toward what industry observers describe as the AI control layer, which includes the orchestration, monitoring, evaluation, governance, permissions, auditability, accountability, escalation paths and human oversight needed to keep AI systems operating safely and effectively.
"The hardest part of enterprise AI is no longer just choosing or building the model," said Judah Phillips, chief AI and product officer at Market Holdings/Squark AI. "The real challenge is everything that happens after the model enters the business -- what data it touches, what actions it can take, who is accountable when it is wrong and how the organization proves it created value."
It's all about control
On the road to greater autonomy, AI systems must earn increasing levels of trust. "You start with assistance, move to recommendations, then approved actions, and only then to bounded autonomy where the system has demonstrated reliability, accountability and economic value," Phillips explained.
When AI is deployed in business processes, the challenge shifts from generating outputs to managing decisions, workflows and accountability. The question is no longer whether a model can produce an answer, but whether businesses can understand, govern and control AI's actions across systems.
Creating a human approval factory is an expensive bottleneck where humans spend time reviewing, escalating and correcting agents.
Tori PaulmanVP analyst, Gartner
Many businesses underestimate the amount of workflow redesign required before autonomous systems can operate with minimal human involvement, said Tori Paulman, VP analyst at Gartner. Regulatory requirements, decision-making structures and exception-handling processes make it difficult to remove humans entirely from the loop.
"Creating a human approval factory is an expensive bottleneck where humans spend time reviewing, escalating and correcting agents instead of doing higher-value work," Paulman said. As a result, businesses can struggle to realize the full value of autonomous systems.
As AI systems interact with multiple applications, tools and data sources, governance extends beyond model performance to visibility, accountability and control. "To unlock real value," Paulman advised, "enterprises must decompose workflows, set decision rights, define supervisor roles, and make agents visible and auditable in systems of record."
The control layer is critical to the successful management of AI deployments.
Infrastructure is key to performance and accountability
While businesses have spent years focusing on AI model capabilities, the model "was never the hard part," said Lexi Reese, CEO and co-founder of AI workflow platform provider Lanai and former Google vice president. "The hard part is the moment AI touches a real organization, and most organizations are completely unprepared for that moment."
Companies can struggle with the management and operational challenges that emerge once AI enters production environments, Reese said. "The models work, it's the management that doesn't," she explained. "We are past the question 'Can AI work?' The real questions now are: 'work for whom, on what and how do you know?'"
Many businesses lack the infrastructure needed to understand what AI systems are doing once they move into production and how those activities contribute to business outcomes, Reese said. "Companies are using AI like labor but accounting for it like software," she noted.
The P&L has no line for AI labor. The org chart has no box for an agent. The workforce plan has no column for supervised machine labor.
Lexi ReeseCEO and co-founder, Lanai
Businesses might know how much they're spending on AI tools, yet, they often lack clear mechanisms to determine what those systems are producing, who owns the outcomes and whether the technology is creating measurable business value. "The P&L [profit and loss statement] has no line for AI labor," Reese said. "The org chart has no box for an agent. The workforce plan has no column for supervised machine labor."
Businesses find themselves struggling to answer basic questions about AI performance and ownership once systems are deployed. "The real enterprise AI problem in 2026 is not model performance," Reese said. "It is the missing infrastructure that tells you what your AI workforce is doing, who owns it and whether it's actually creating value."
Business and technology leaders are finding that successful AI deployments depend as much on organizational processes and communication as they do on the underlying technology. "The biggest ongoing misconception that we see is that AI can work around organizational barriers," said Randall Hunt, CTO at cloud native services provider Caylent. "Most of the issues we face in our implementations are communication or process barriers more than technical barriers."
Before deploying AI models, establish clear ownership, redesign processes, define decision rights and build the controls necessary to safely scale autonomous systems. "The winners," Phillips said, "will be the organizations that treat AI less like a tool and more like an operating capability."
Liz Hughes is an award-winning editor and writer covering AI and emerging technology and the former editor of AI Business and IoT World Today.
