For CIOs, this creates a practical sequencing issue. Before scaling agents, teams need a service map of where decisions are made, where exceptions occur, which systems are authoritative and where human approval is still mandatory. If those basics are unclear, autonomy will amplify hidden operating weaknesses faster than it creates efficiency.
A useful decision test is to classify candidate use cases into three tiers: advisory tasks, constrained execution tasks and autonomous decision tasks. Many enterprises are trying to jump too quickly into the third category when the first two would generate cleaner evidence, safer adoption and better governance patterns. The point is not to slow innovation unnecessarily, but to avoid confusing fluency with operational readiness.
There is also an architectural implication: observability matters as much as intelligence. Enterprises will need logs, validation checkpoints, rollback mechanisms and clear escalation paths so they can explain why an agent acted, not merely what it produced. Without that instrumentation, auditability, incident response and business trust will lag behind deployment speed.
In short, the winners are unlikely to be the organizations with the most ambitious demos, but those that redesign workflows so autonomy is introduced where control, measurement and accountability already exist.
It’s a common belief that AI agents can automate almost anything. But that unrealistic assumption can set enterprises on equally unrealistic trajectories; Gartner predicts that 40% of agentic AI projects will collapse by next year. The prevailing consensus is that the model is almost never why these projects fail.
AI projects implode when teams move straight into building agents without determining what success looks like — or what happens when things go wrong at scale, said Rohit Poduval, a trust and safety engineering leader who works at a major retailer and is a member of the global think tank Integrity Institute, where he has co-authored policy responses on AI safety and children’s online privacy.
And things do go wrong quite often. “The corrective path is to start with the process, not the agent,” said Medhat Galal, senior vice president of engineering at Appian Corp., a provider of AI process automation.
Where AI agents go wrong
“In most of the failures we see, the technology did exactly what it was asked to do; the trouble is what it was asked to do,” said Justin Bolles, CTO at Resultant, a data, technology and AI consulting firm.
Putting agents to work in existing processes and expecting equivalent — or better —outcomes than employees can produce is an unrealistic and largely unsuccessful effort.
Most enterprise workflows were never designed for machine execution, so they typically “involve undocumented workarounds and tribal knowledge,” explained Priya Sawant, senior vice president of engineering at ASAPP, a provider of AI agents for enterprise contact centers. “Deploying an agent into that environment doesn’t fix the mess; it makes it fail faster and at scale.”
In other words, an agent’s work can go awry when the process isn’t clear or the data supporting it is poor or incomplete. Many agentic projects fail because “the workflow being automated was never as clean as the project team thought it was,” explained Rishi Bhargava, co-founder at Descope, the maker of a customer and agent authentication platform.
As Bhargava described it, a new agent will simultaneously come up against issues like ambiguous ownership, undocumented exceptions and steps that depend on someone’s institutional knowledge — with potentially disastrous results.
Even if an agent can navigate a workflow well enough to produce an outcome, that outcome may not be the one it was directed to produce. Most enterprise operating models “treat AI as a transactional endpoint,” said Sekhar Sarukkai, co-founder and CEO of ChatSee.ai. According to Sarukkai, when a user submits a request and the model returns a plausible answer, that interaction is considered successful. But plausible is not the same as accurate.
“That model breaks down as agents begin interpreting intent, using tools and taking actions across workflows,” Sarukkai added. He cited as examples:
-
A customer service agent that responds fluently but fails to escalate the ticket;
-
A finance agent that correctly extracts information but applies the wrong exception policy; or
-
A coding agent that generates valid code while modifying the wrong repository.
“In each case, the technology appears to be functioning, but the business outcome is wrong,” Sarukkai said.
Finding fixes to agentic process management
This doesn’t mean that agentic AI has no application, just that it can’t be deployed within unprepared systems. Some processes may need to be redesigned, others may need one or more parts extracted to better clarify the agent’s mission, while others should be ditched or replaced entirely.
“Companies need to define the work, integrate the data and systems around it, establish guardrails and decision rights, and then introduce autonomy in controlled stages,” Galal said.
Without the proper controls in place at every stage, agents will quite literally run with what they have — and run over what they don’t. Agents “don’t fill in the gaps,” Poduva said. If you haven’t explicitly told them what to do in a given scenario, “they’ll either hallucinate an answer or do something unpredictable,” he added.
Providing proper controls means developing more than policies and a few rules for agents to follow.
The phrase ‘we have guardrails’ is “the most dangerous sentence” in enterprise AI, according to Raj Koneru, founder and CEO of Kore.ai, an enterprise AI platform and agentic AI company. “What is needed is a foundation that addresses the illusion of governance and provides real control,” Koneru added. At the root of this lies an age-old wisdom for keeping business on track: the KISS (keep it simple, stupid) system, which works very well in successfully using agents. Unfortunately, teams tend to point agents at “the impressive, judgment-heavy problem that demos well,” instead of the “boring, high-volume, well-bounded process” where agents “actually pay off,” said Dr. Daniel Tiarks, co-founder and CTO at Cambrion, an agentic AI data processing platform provider.
Tiarks suggested the following ways to control and benefit from agentic AI through better process management:
-
Start with a single bounded, high-volume process that has a clear right answer.
-
Wrap the model in deterministic validation and keep a human gating the edge cases.
-
Define the accuracy and throughput number you need before you start, then measure against it.
-
Demand traceability. Every output should trace back to its source for audit and legal defensibility.
-
Treat the model as swappable infrastructure, i.e., model-agnostic, API/model context protocol into the existing stack, not a one-time bet on a single vendor.
-
Redesign the process around where the agent is reliable; don’t bolt an agent onto a broken workflow.
Remember that agent models rarely fail in isolation; they fail inside processes that were never designed for autonomous action.
“The fix isn’t better models,” said Kristof Horompoly, head of AI at ValidMind, an AI governance platform. “It’s narrowing scope to bounded, well-instrumented tasks, building real evaluation harnesses before you scale, and treating these as operational redesigns rather than technology projects.”
Has your AI agent project failed — or succeeded against the odds? Email your story to [email protected].
Enjoyed this article? Sign up for our newsletter to receive regular insights and stay connected.