Most companies that launch AI chatbots for customer service end up pulling them.
In 2025, Sinch surveyed more than 2,500 business leaders across 10 countries and found that 74% of companies that deployed AI-powered customer interaction projects had to shut them down or roll them back. That number alone is striking, but the stranger finding is underneath it. Companies with the most mature governance frameworks had an 81% rollback rate, seven points worse than average. Organizations with the most mature governance were pulling the plug more often than those with less developed oversight.
Sinch is calling it the "AI Production Paradox." The company's CPO, Daniel Morris, said the industry has long assumed that better governance leads to better outcomes, but the data tells a different story.
There's a generous way to read this. Greg Carlucci, senior director analyst at Gartner, told CX Dive that the high rollback rate isn't necessarily a bad sign. "The AI movement is moving so quickly that there inevitably will be a lot of tests and learning," he said. Companies with real monitoring in place are simply more likely to catch a problem before customers start posting about it. A company with no monitoring wouldn't know its chatbot was leaking data until someone told them.
That reframe is reasonable. But the operational cost of all this detecting and rolling back is adding up fast. Sinch found that 84% of AI engineering teams working on customer communications are spending at least half their time building guardrails and safety controls instead of actual product features. That's a lot of engineering talent going toward making sure nothing goes wrong, and most of these agents are still getting pulled.
The confidence gap tells its own story. 90% of leaders described themselves as ready for AI before they launched. Among those already in production, 75% had experienced at least one governance rollback. And inside the same companies, technical leaders reported rollback rates of 77% while business leaders said 69%. The people closest to the code consistently knew the situation was worse than the people signing the checks.
When Sinch looked at what actually predicts whether a deployment survives, the answer was infrastructure satisfaction, meaning how capable and reliable the underlying communications platform was before AI got layered on top. That had the strongest correlation of anything they tested across thousands of variable pairs, stronger than investment level, AI maturity, or how sophisticated a company's guardrails were. Sinch sells communications infrastructure so the finding points conveniently toward their own product category, but it lines up with what Kristina McElheran at the University of Toronto found in completely separate research on AI adoption in manufacturing. Older firms that adopted AI actually saw their structured management practices deteriorate after deploying it, and that deterioration alone explained nearly a third of their productivity losses. "AI isn't plug-and-play," McElheran said. "It requires systemic change, and that process introduces friction, particularly for established firms."
Both findings point in the same direction. Companies are bolting autonomous AI onto workflows and infrastructure built for human operators and discovering the foundation wasn't designed for what's now sitting on top of it.
Nobody is giving up. Brian Weber, a VP analyst at Gartner, said "an agentless contact center is not yet technically feasible, nor is it operationally desirable." Analysts suggest the companies pulling their chatbots are fixing what broke and redeploying. As Carlucci put it: "Until you actually see it in use with the customer, there are things that you'll need to adjust."
The 81% rollback rate among the best-prepared teams is the number that should bother people. The entire industry spent the last two years building governance on top of infrastructure that wasn't designed for it, and the result is that safety controls are generating overhead instead of stability. The organizations that pull ahead will be the ones that stop layering rules on top and start rebuilding what's underneath.
