AI Agents Succeed But Processes Fail

Gartner predicts that 40% of agentic AI projects will collapse by next year. The prevailing consensus among experts is that the underlying model is almost never the reason these initiatives fail. Instead, the failure lies in the foundation of the business processes themselves. When teams move straight into building agents without determining what success looks like, or what happens when things go wrong at scale, the project is likely doomed.

“The corrective path is to start with the process, not the agent,” said Medhat Galal, senior vice president of engineering at Appian Corp.

Rohit Poduval, a trust and safety engineering leader at a major retailer, noted that teams often bypass the necessary planning stages. He explained that projects implode when organizations fail to anticipate failure modes before deployment.

Process Over Technology

In many observed failures, the technology functioned exactly as programmed. The issue was the instruction set it received. Justin Bolles, CTO at Resultant, pointed out that the trouble usually lies in what the agent is asked to do, rather than how it executes the code.

Most enterprise workflows were never designed for machine execution. They typically involve undocumented workarounds and tribal knowledge. Priya Sawant, senior vice president of engineering at ASAPP, stated that deploying an agent into a messy environment does not fix the problem; it simply causes the system to fail faster and at scale.

Rishi Bhargava, co-founder at Descope, explained that workflows are rarely as clean as project teams assume. A new agent often encounters ambiguous ownership, undocumented exceptions, and steps that rely entirely on institutional knowledge.

The Accuracy Gap

Even if an agent handles a workflow well enough to produce an outcome, that outcome may not be the one intended. Most enterprise operating models treat AI as a transactional endpoint. According to Sekhar Sarukkai, co-founder and CEO of ChatSee.ai, a plausible answer is often mistaken for an accurate one.

This model breaks down as agents begin interpreting intent and taking actions across workflows. Sarukkai cited examples where a customer service agent responds fluently but fails to escalate a ticket, or a finance agent extracts information correctly but applies the wrong exception policy.

“In each case, the technology appears to be functioning, but the business outcome is wrong,” Sarukkai said.

Establishing Real Control

Agentic AI has applications, but it cannot be deployed within unprepared systems. Companies need to define the work, integrate the data, establish guardrails, and introduce autonomy in controlled stages.

Without proper controls, agents will run with what they have and run over what they don’t.