Three-Quarters of Enterprise AI Customer Service Deployments Are Rolled Back, New Study Finds
A Sinch survey of 2,527 AI decision-makers finds 74% have shut down live AI customer communications agents, with rollback rates rising to 81% among the most governance-mature organizations.
Overview
Nearly three-quarters of enterprises that have deployed AI customer communications agents have been forced to shut them down or roll them back — and the problem is worse, not better, for organizations with the most sophisticated safety controls. That is the central finding of The AI Production Paradox, a global study published on May 13 by Swedish communications-as-a-service firm Sinch, which surveyed 2,527 senior decision-makers across ten countries and six industries between January and February 2026.
What the Study Found
The headline figure — a 74% rollback rate — already challenges the narrative that enterprise AI adoption is proceeding smoothly. More striking is what happens when organizations invest heavily in governance: the rollback rate climbs to 81% among organizations that describe their guardrails as “fully mature,” according to Sinch’s published findings.
The apparent paradox dissolves under scrutiny. Daniel Morris, Chief Product Officer at Sinch, argues that maturity in monitoring — not failure in execution — drives the elevated number among experienced organizations. “The most advanced organisations aren’t failing less; they’re seeing failures sooner,” Morris told Capacity. “Higher rollback rates reflect better monitoring and control, not weaker performance.”
Morris was more pointed in a statement published in the official Sinch research blog: “If governance was the fix, the most mature teams would roll back less. They don’t. What’s breaking isn’t the policy layer. It’s reliability in the real system: data, workflows, integrations, and edge cases. And the cost is real: 84% of AI engineering teams spend at least half their time on safety infrastructure instead of improving the customer experience. That’s the guardrail tax.”
Why Agents Fail in Production
The study’s most consequential finding for practitioners may be the breakdown of rollback causes. According to Sinch’s report chapter on AI production challenges, the leading cause is PII or customer data exposure, cited by 31% of respondents who experienced a rollback. Hallucination or brand risk follows at 22%, and lack of auditability — the inability to diagnose what went wrong — accounts for 16%.
The downstream effects of these failures extend beyond the individual failed deployment. Among organizations that rolled back an AI agent, 35% reported a surge in human support agent load, and 34% cited reputational damage and loss of customer trust as a primary consequence, per Sinch’s blog findings.
The Infrastructure Gap
The research frames inadequate infrastructure — not governance frameworks or investment levels — as the strongest predictor of deployment failure. Sinch found a 0.52 correlation between infrastructure satisfaction and AI deployment confidence, stronger than either investment level or guardrail maturity. Among the specific gaps identified, 42% of respondents reported that their current provider offered insufficient reliability for AI at scale, 37% cited limited multi-channel capability, and 32% reported a lack of AI platform integrations.
As a workaround, 55% of organizations said they are custom-engineering solutions for preserving customer context across channels, a costly diversion of engineering capacity that compounds over time, according to Sinch.
Anton Efimenko, SVP Software Engineering at Sinch, described the cumulative burden on engineering teams: “Every team needs to decide what controls belong at the platform layer and what their engineers should build on top, because 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.”
Spending Priorities Signal a Structural Mismatch
The study’s spending data underscores the tension between safety and capability development. Three-quarters of respondents — 75% — rank trust, security, and compliance among their top three AI spending priorities. Only 63% say the same of AI development itself, according to The Register’s reporting on the study. This disparity reflects what the Sinch report calls the “guardrail tax”: the engineering overhead of running AI safely at scale exceeds most organizations’ initial projections.
Despite the challenges, investment continues to climb. The study found that 98% of enterprises are increasing their investment in AI communications in 2026, and 88% expect to have a live AI agent in production by year-end, per Sinch.
What We Don’t Know
The study was commissioned by Sinch, a vendor that sells communications infrastructure to enterprises deploying AI. The framing of infrastructure as the primary lever for improvement — rather than model quality, use-case selection, or organizational change management — aligns closely with Sinch’s commercial interests, a limitation worth noting when evaluating the study’s conclusions.
The survey also captures a single cross-sectional moment, covering deployments as of January-February 2026. It does not track whether organizations that rolled back agents later re-deployed successfully, or whether rollback rates vary by use case, channel type, or AI model provider.
Sinch has indicated that full regional and industry-level breakdowns will be released in Q2 2026.