Nearly three-quarters of enterprises have already rolled back or shut down a customer-facing AI agent after deployment. That figure comes from Sinch’s “AI Production Paradox” report, an independent survey of 2,527 senior decision-makers across 10 countries and six industries released May 13.

The headline number is striking. The number underneath it is stranger.

The Governance Paradox

Among organizations with the most mature governance frameworks, the rollback rate climbs to 81%. Companies that invested most heavily in compliance, safety protocols, and oversight are failing more visibly, not less.

“If governance was the fix, the most mature teams would roll back less, not more,” Daniel Morris, CPO at Sinch, told MarTech. “Engineering teams are spending most of their time building and maintaining safety systems instead of focusing on improving the customer experience. That’s the guardrail tax that slows organizations down.”

Sinch frames the higher rollback rate among governed teams as a monitoring advantage: better oversight surfaces failures that less mature organizations simply miss. “The most advanced organizations aren’t failing less; they’re seeing failures sooner,” Morris said in Sinch’s press release. “Higher rollback rates reflect better monitoring and control, not weaker performance.”

Why Agents Fail in Production

The survey data pinpoints specific failure modes. Nearly one-third of organizations cited customer data exposure as the leading cause of rollback, according to CX Dive’s reporting. Another 22% cited hallucination or brand risk. Sixteen percent said the problem was simpler: they couldn’t diagnose what went wrong.

Greg Carlucci, senior director analyst at Gartner, told CX Dive that rollbacks reflect learning, not failure. “The AI movement is moving so quickly that there inevitably will be a lot of tests and learning, so I don’t necessarily feel that the high level of rollback is a negative,” he said. “In fact it’s almost, in my view, a positive, because the organizations that are launching these tools want to make sure that they’re getting it right.”

But the preparation gap is real. Brands can run thorough internal testing, pilot likely customer questions, and stress-test edge cases. “But it isn’t a perfect science,” Carlucci said. “Until you actually see it in use with the customer, there are things that you’ll need to adjust, whether it’s the data that the AI agents trained on, the tools that it has access to.”

Chuck Gahun, principal analyst at Forrester, pointed to cascading incorrect responses, data quality issues, and lack of tracing and logging at each step of the agent’s workflow as common culprits, according to CX Dive.

The Guardrail Tax

The survey’s cost data adds another layer. Enterprises invest more in trust, security, and compliance (76% of AI budgets) than in AI development itself (63%), according to Sinch’s findings. A full 84% of AI engineering teams spend at least half their time on safety infrastructure rather than building the product.

Jayashree Iyangar, global lead of CX data and AI at HGS, told MarTech that the real challenge lies in operations, not deployment. “The key question is how AI can be orchestrated seamlessly across multiple channels, not whether it can be deployed in one,” she said. Risk varies by use case: a marketing chatbot that fumbles a promotion carries less weight than a service agent that mishandles a billing dispute. “Human-in-the-loop oversight remains central in service environments where the risk of negative customer impact is higher.”

Deployment Isn’t Slowing Down

Despite the rollback rate, the market is accelerating. Sinch found that 62% of enterprises already have AI agents live in production. Eighty-eight percent expect to have agents deployed within a year. And 98% are increasing AI investment in 2026.

The disconnect between rollback rates and investment rates tells a specific story: enterprises are not abandoning agents. They are discovering that production is where the real engineering begins, and that governance frameworks, as currently built, create friction faster than they prevent failure.