Three out of four enterprises have pulled an AI agent out of production after it went live. That number comes from a Sinch survey of 2,500 senior decision-makers across ten countries. It is not a small sample. It is not a early-adopter curiosity. It is the current state of enterprise AI.
But here is the part nobody is talking about. The companies with the most mature governance frameworks rolled back agents at an even higher rate: 81 percent. The best-governed organizations are killing more AI agents than anyone else.
That sounds like failure. It is the opposite.
The Counterintuitive Math
Sinch's CPO Daniel Morris put it bluntly: the most advanced organizations are not failing less. They are seeing failures sooner. Higher rollback rates reflect better monitoring and control, not weaker performance.
Think about what that means. A company with no governance deploys an AI agent and it runs for months, quietly making bad decisions nobody catches. A company with mature governance deploys the same agent and shuts it down in a week because their monitoring flagged a data exposure risk on day three.
The first company reports zero rollbacks. The second reports one. Which one is actually safer?
This is the trap in how we talk about AI agent success. We count deployments and rollbacks as if they are the same metric across every organization. They are not. A company that never rolls back is not necessarily running better agents. It might just not be watching.
Gartner's Greg Carlucci framed it the same way in a recent CX Dive interview: the high rollback rate is almost a positive, because organizations launching these tools want to make sure they are getting it right.
What Actually Goes Wrong
The Sinch data breaks down the rollback triggers, and they are not what most people assume.
Customer data exposure is the leading cause, cited by 33 percent of organizations. Not hallucinations. Not latency. Not cost. The agent got access to customer data it should not have had, and someone caught it.
Hallucination and brand risk came second at 22 percent. The agent said something wrong, or said something that did not represent the brand, and it went public.
Third: the inability to diagnose what went wrong. Sixteen percent of organizations rolled back an agent because they could not figure out why it was making bad decisions. They saw the symptoms. They could not trace the cause.
That third one should scare you more than the first two. Data exposure and hallucination are known risks with known solutions: access controls, output filtering, brand guardrails. The inability to diagnose means the agent is a black box that nobody can debug. You cannot fix what you cannot trace.
Forrester's Chuck Gahun pointed to the core issue: cascading incorrect responses, data problems, and lack of tracing and logging at each step of the agent workflow. The agent is only as good as the data it is trained on and the permissions it is granted. And most brands do not have unified, clean data.
The Data Problem Nobody Solves First
Carlucci made the point that should be obvious but apparently is not: an AI agent needs a large amount of unified, organized, centralized data to function. That takes time to build. It takes investment. Only a few companies are ready for fully functional autonomous AI agents.
This is where production deployment falls apart for most organizations. The pilot works because the data is clean. The pilot is a controlled environment with curated inputs and a small team watching every output. Production is the opposite: messy data, legacy systems, real customers with edge cases, and nobody watching.
The agent does not fail because the model is bad. It fails because the data layer was never built. The organization skipped the boring infrastructure work and went straight to the agent demo.
This is the same pattern that Gartner flagged when they predicted 40 percent of agentic AI projects will be canceled by 2027. The failures are management problems, not model problems. Poor governance, undefined business value, insufficient operational discipline. The technology is fine. The humans around it are not.
What the Rollback Rate Actually Measures
Here is the reframe that matters for marketing teams and CMOs.
The rollback rate is not a failure metric. It is a maturity metric. A company that rolls back agents quickly is a company that has the infrastructure to detect problems, the authority to act on them, and the culture to admit something is not working.
A company that reports zero rollbacks either has perfect agents (unlikely) or has no visibility into what their agents are doing (probable). The governance gap most organizations face is not about having policies on paper. It is about having the real-time monitoring, the audit trails, and the decision authority to pull the plug when something goes wrong.
The Sinch survey confirms this. Sixty-two percent of enterprises already have AI agents in production. Nearly nine in ten say their AI agents will be live within a year. The money is committed. The deployments are happening. The question is whether organizations can see what their agents are doing in real time and act on it.
The rollback rate tells you whether the answer is yes.
The Industry Split
The rollback numbers are not uniform. Sinch found rates ranging from 66 percent in the technology sector to 85 percent in professional services. That gap tells a story.
Technology companies have more infrastructure for monitoring and debugging. They are more likely to catch problems early, roll back fast, and redeploy with fixes. Their lower rollback rate does not mean fewer problems. It means faster cycles of deploy, catch, fix, redeploy.
Professional services firms have higher rollback rates because their governance frameworks are catching more issues before those issues become public. They have more at stake: client confidentiality, regulated advice, reputation. Their 85 percent rollback rate is not a failure of AI. It is the governance system working as designed.
For marketing teams, the lesson is specific. Your AI agents touch customer data, brand voice, and public communications. The rollback rate you should be tracking is not whether agents get deployed. It is how fast you catch them when they go wrong. Agent abandonment is not a sign of failure when it happens because your monitoring caught a real problem.
The McKinsey Number Nobody Mentions
McKinsey found that 80 percent of organizations have already encountered risky agent behaviors, including unauthorized data exposure and improper transactions. That is not 80 percent of pilot programs. That is 80 percent of all organizations deploying agents.
The gap between 80 percent encountering risky behavior and 74 percent rolling back agents means there is a meaningful chunk of organizations that know their agents are doing something risky and have not pulled them yet. They are running agents they know are problematic because they do not have the authority, the infrastructure, or the will to shut them down.
That is the real governance gap. Not the absence of policies. The absence of operational nerve.
What Changes in the Next Year
Sinch says nearly 90 percent of enterprises will have AI agents live within a year. Gartner says 40 percent of those projects will be canceled by 2027. Both of these things can be true at the same time, and they probably are.
The organizations that will succeed are the ones that treat rollback as a feature, not a bug. They deploy fast, monitor aggressively, and shut down without hesitation when the data says something is wrong. They do not celebrate deployment counts. They celebrate detection speed.
The organizations that will fail are the ones that treat rollout as a one-way door. They deploy an agent, declare success, and resist pulling it back because it looks like an admission of failure. Meanwhile the agent is quietly making decisions nobody audits, exposing data nobody tracks, and eroding customer trust nobody measures.
The rollback rate is the most honest metric in enterprise AI right now. If yours is zero, the question is not whether your agents are perfect. The question is whether you would know if they were not.