rewrite this title New Sinch Data Reveals 74% of Enterprises Have Rolled Back AI Agents – UC Today

The AI industry has spent years fixated on one problem: getting AI out of the lab and into production.

According to new research from cloud communications vendor Sinch, that battle is largely won – but a bigger one has taken its place.

Sinch’s new report, The AI Production Paradox, is based on an independent survey of 2,527 senior decision makers across 10 countries and six industries, and paints a picture of an enterprise AI market that has scaled rapidly but is struggling to sustain what it has built.

The report claims that 74 percent of enterprises have already rolled back or shut down a live AI customer communications agent following deployment – suggesting that for many organisations, going live was the easy part.

“The industry has assumed that better governance leads to better outcomes. But that’s not enough,” said Daniel Morris, CPO at Sinch.

“If governance was the fix, the most mature teams would roll back less, not more.”

Deployment Isn’t The Problem Anymore

The survey finds that 62 percent of enterprises already have AI agents live in customer communications – a figure that pushes back against the narrative that the enterprise market is stuck in endless pilot phases.

The challenge, Sinch argues, has fundamentally shifted. Getting AI into production is no longer the primary barrier. What happens next is.

That shift has significant implications for how enterprises think about AI investment and infrastructure.

Many organisations built their way into production without the underlying systems needed to maintain performance, reliability and control at scale. Now, according to Sinch, they’re paying the price.

The scale of rollbacks is notable across the board, but particularly so among the organisations best positioned to avoid them.

Among enterprises with the most mature AI governance frameworks, the rollback rate reportedly climbs to 81 percent – higher than the 74 percent overall average.

Sinch’s interpretation is that mature monitoring capabilities allow these teams to identify failures that less sophisticated organisations are simply missing.

“The most advanced organisations aren’t failing less; they’re seeing failures sooner,” Morris said. “Higher rollback rates reflect better monitoring and control, not weaker performance.”

Governance Investment Alone Isn’t Solving It

The data suggests enterprises are not ignoring the problem.

Investment in trust, security and compliance (76 percent) now reportedly outpaces spending on AI development itself (63 percent), making it the single largest investment category in enterprise AI programmes.

This is where Sinch introduces the concept of the “Guardrail Tax” – the idea that safety infrastructure has become a significant and growing drain on engineering capacity. 84 percent of AI engineering teams reportedly spend at least half their time on safety systems rather than building new features or improving customer experience.

For organisations under pressure to demonstrate AI ROI, that’s a compounding cost with no obvious end point.

Sinch’s data identifies communications infrastructure satisfaction as the strongest predictor of successful AI deployment – stronger than governance maturity or overall investment levels. That conclusion conveniently aligns with Sinch’s own product offering.

More than half of enterprises (55 percent) say they are building custom infrastructure simply to manage cross-channel context, and 86 percent have evaluated or are actively considering switching communications providers.

The Stakes Keep Rising

Despite the scale of rollbacks and the engineering burden they represent, appetite for AI investment shows no signs of slowing. 98 percent of enterprises report they are increasing AI communications spend in 2026 – meaning the gap between ambition and reliable execution is set to widen further before it narrows.

“Engineering teams are spending most of their time building and maintaining safety systems – a lot of which their communications infrastructure should be providing,” Morris added. “That’s the guardrail tax that slows organisations down.”

The AI Production Paradox early access report is available now, with full regional and industry breakdowns expected before the end of June.