AI customer service rollbacks hit 74%

74% is the number that cuts through the AI-customer-service hype. Sinch, a Swedish communications software company, said on May 13 that 74% of enterprises had already rolled back or shut down an AI customer-communications agent after deployment because of a governance failure. The company said the figure came from a global survey of more than 2,500 AI decision-makers in its report, *The AI Production Paradox*. (theregister.com) The detail that matters most is what happened at the companies that were supposed to be better prepared. Sinch said rollback rates rose to 81% at organizations with “fully mature guardrails.” Daniel Morris, Sinch’s chief product officer, said that suggested the problem was not simply a lack of policy, because “the most advanced organizations aren’t failing less; they’re seeing failures sooner,” with higher rollback rates reflecting stronger monitoring and control. (theregister.com) That changes the frame of the story. If companies with more mature controls are still pulling systems out of production, the issue looks less like model capability alone and more like what happens after a bot goes live: who owns it, what data it can access, how it is monitored, and how quickly someone can intervene when it goes wrong. That is an inference from Sinch’s findings, not a direct quote from the company, but it is supported by the pattern in the data it published. (theregister.com) Sinch’s own numbers point in that direction. The company said 84% of AI engineering teams spend at least half their time on safety infrastructure, and 75% of respondents ranked trust, security and compliance in their top three priorities, ahead of AI development itself at 63%. Sinch said the rollback rate was consistent across regions and industries, which it said meant company size and budget were not meaningful protective factors. (theregister.com) That makes this a management story as much as a technology story. A customer-service bot does not fail only when it gives a wrong answer. It can also fail when it cannot reliably pull the right account data, cannot hand off to a human cleanly, cannot document what it did, or cannot stay inside policy limits set by legal, compliance or security teams. Sinch’s description of “governance failures” does not break out each cause in the excerpts available, but its emphasis on production management and safe operation points to operational controls as a central bottleneck. (theregister.com) That is why finance, audit and risk teams are showing up more often in AI conversations. WBUR’s *Here & Now* reported on May 13 that financial analysts were urging consumers not to share sensitive personal-finance information with chatbots, highlighting a broader concern about what data these systems receive and how it is handled. Accountancy Age wrote on May 14 that accountants in 2026 were facing growing liability as their advisory role expanded into areas including data security and compliance, describing a profession dealing with blurred boundaries and control risks. (wbur.org) (accountancyage.com) Those are different contexts, but they point to the same practical question inside companies: when an AI system touches customer interactions, who is accountable for the controls around it? The rollback statistic does not mean AI in customer service is disappearing. Sinch itself sells AI-related customer communications products and infrastructure, and it has continued to launch new AI offerings this year, including voice and agentic-conversation tools. What the survey does suggest is that many enterprises are discovering that deployment is the easy part compared with running these systems safely in production. (sinch.com) The next useful data point will be whether companies respond by narrowing AI use cases, rebuilding governance around them, or restoring more human agents alongside automation. Gartner had already reported in 2025 that many customer-service leaders planned to retain human agents rather than move to fully agentless models, a direction that fits the rollback pattern now being reported. (theregister.com)