Sinch Finds 73% Rolled Back AI Customer Agents as Governance Realities Hit Home

Nearly three in four enterprises (73%) have been forced to shut down or roll back a live AI customer communications agent following a "governance failure". The figure, drawn from Sinch's new global research report, The AI Production Paradox, does not necessarily mean you should avoid adopting AI customer agents altogether. The study, which surveyed 2,527 senior decision-makers across ten countries and six industries, is instead a wake-up call to the realities of governance expectations and infrastructure requirements for CX AI in 2026.

The report found that for organisations with the most advanced governance frameworks, the roll-back rate rises to 81%, several points above the overall average. This may seem like governance itself is to blame but it simply means that more sophisticated organisations are monitoring more carefully, detecting failures earlier, and acting on them. Organisations without this level of governance oversight may just not realise their AI is underperforming.

Daniel Morris, CPO at Sinch, explains: "The most advanced organizations aren't failing less; they're seeing failures sooner. Higher rollback rates reflect better monitoring and control, not weaker performance."

The strain is real, however. The same data shows that 84% of AI engineering teams spend at least half their time building and maintaining safety infrastructure rather than improving the customer experience itself. Sinch describes this as a "guardrail tax", which represents one of the hidden costs that accumulate once AI moves into production, and teams need to start accounting for it in advance.

This guardrail tax may also be a symptom of a problem one layer below. Sinch found that more than half of enterprises (55%) are custom-building the ability to preserve customer context when someone switches channels because their infrastructure does not handle it natively. Anton Efimenko, SVP of Software Engineering at Sinch, describes how the development work can add up: "The cost of building custom guardrails compounds over time, especially as the team moves through the product lifecycle. Each new agent, each new channel, each new deployment adds to the pile. And eventually you lose that momentum when it comes to outperforming on the market."

The Sinch findings are consistent with what other research is showing. Gartner, for example, found that only 13% of organisations feel prepared for AI governance at scale. This helps explain why so many CX AI projects are failing. A 2025 enterprise architecture study published by Cornell University, examining GenAI adoption across large organisations, identified governance gaps, poor data maturity, and organisational complexity as among the most significant barriers to sustainable deployment at enterprise scale.

Despite governance friction, Sinch's research also uncovered that 98% of enterprises are increasing investment in AI communications. Broader signals from across the CX AI market point in the same direction, with major vendors including Salesforce, ServiceNow, Zendesk, Five9, and 8x8 all reporting strong AI-related momentum across their respective earnings reports.

Most enterprises pulling back from specific deployments are not abandoning AI. They are re-evaluating, reiterating, and redeploying. Viewed through this lens, the rollbacks are less a sign of collapse than the market entering a more mature phase.

The early narrative around AI in customer experience was built on straightforward promises to connect a model, reduce handle times, automate volume, and cut cost. What enterprises are now discovering is that production-grade CX AI requires governance, escalation logic, hallucination controls, compliance infrastructure, auditability, and continuous monitoring. Those requirements were either underestimated or undisclosed in many early initiatives.

Governance is not the problem; we are just getting a clearer picture of the guardrails required. For some, it may indeed prove more cost-effective to buy an off-the-shelf solution from an AI vendor versus building their own custom agent. More broadly, the report could also indicate that AI hype is starting to settle down as a more sober view of deployments emerges.