May 14, 2026
Are Enterprises Deploying AI Agents Faster Than Their Governance Can Handle?
A large majority of enterprises have rolled back or completely shut down an AI agent in customer communications after it went live, according to a new global study from Sinch.
Most of the industry still treats deployment as the biggest challenge, but the real problems start after the AI agent is already live. In reality, most organisations have already crossed that line and are now struggling with what comes next.
Sinch’s research report, The AI Production Paradox, found that 62% of enterprises already have AI agents operating in live customer communications environments. Yet 74% of those organisations have had to reverse a deployment because of a governance failure, raising questions about whether the systems and infrastructure behind these agents are keeping up with the pace of rollout.
More Governance, More Rollbacks
The rollback rate climbs even higher among organisations that have invested the most in oversight. Enterprises with the most developed governance frameworks reported an 81% rollback rate, above the overall average. This is contrary to the expectation that stronger guardrails should produce fewer problems. According to Daniel Morris, CPO at Sinch, the explanation is that better-prepared organisations are not necessarily failing more often, but are detecting failures sooner because their monitoring systems are more advanced.
“The industry has assumed that better governance leads to better outcomes. But that’s not enough,” Morris said. “Engineering teams are spending most of their time building and maintaining safety systems, a lot of which their communications infrastructure should be providing, instead of focusing on improving the customer experience. That’s the guardrail tax that slows organisations down.”
Spending on Trust Exceeds Spending on AI Development
Enterprises invest more in trust, security, and compliance (76%) than they do in AI development itself (63%), making safety-related spending the top investment category in enterprise AI communications programmes. Meanwhile, 84% of AI engineering teams report spending at least half their working hours on safety infrastructure rather than on the products and features they are supposed to be building.
When engineering teams spend most of their time building safety systems that should already exist in the communications platform, the result is fewer people working on what AI deployment was meant to improve in the first place, which is the customer experience itself.
Infrastructure Emerges as the Strongest Predictor of Success
According to the report, the best predictor of whether an AI deployment will success is governance maturity. The satisfaction with the underlying communications infrastructure also plays a significant role.
Eighty-seven percent of organisations said high-performance infrastructure is essential or very important to successful AI deployment. And yet most organisations reported that their current provider fails to meet expectations in at least one meaningful area. More than half of enterprises have had to build custom infrastructure themselves just to manage cross-channel context, which is a fundamental requirement for any AI agent expected to maintain a coherent conversation across messaging, voice, and email.
The custom-build burden feeds directly into the guardrail tax Morris pointed out, that is, engineering teams spending their time piecing together platform-level capabilities instead of developing better AI experiences.
Enterprises Are Looking for New Communications Partners
Dissatisfaction with existing infrastructure is already producing commercial consequences, which is pushing 86% of enterprises to either evaluate or are already actively evaluating new communications providers. The search for new partners comes as AI adoption in contact centres accelerates and organisations discover that existing platforms cannot support the demands AI agents place on them.
Despite the high rollback rates, there is no indication that enterprises plan to slow down. Ninety-eight percent said they will increase their AI communications investment in 2026, suggesting that the industry sees these failures as engineering and infrastructure problems to solve, not reasons to retreat.
