The AI industry has spent years grappling with one problem: getting AI out of the lab and into production.
Most of these battles have been won, but bigger ones are emerging, according to new research from cloud communications provider Sinch.
Shinchi’s new report; AI Production Paradoxis based on an independent survey of 2,527 senior decision-makers across 10 countries and 6 industries, and paints a picture of an enterprise AI market that has grown rapidly but is struggling to maintain what it has built.
According to the report, 74% of companies have already deployed and then rolled back or terminated live AI customer communication agents. This suggests that for many organizations, going live is the easy part.
“The industry has assumed that better governance leads to better outcomes, but that is not enough,” he said. Daniel Morris, CPO at Sinch.
“If governance is the solution, the most mature teams will make fewer rollbacks, not more.”
Distribution is no longer an issue.
The survey found that 62% of companies already have AI agents in their customer communications. These numbers run counter to the narrative that the enterprise market is stuck in an endless pilot phase.
Sinch argues that the problem has fundamentally changed. Putting AI into production is no longer a major barrier. Here’s what happens next:
These changes have significant implications for the way companies think about AI investments and infrastructure.
Many organizations have built production environments without the underlying systems needed to maintain performance, reliability, and control at scale. Now, according to Sinch, they are paying the price.
The scale of rollbacks is notable overall, but especially for organizations that are best positioned to prevent them.
Companies with the most mature AI governance frameworks reportedly saw their rollback rate increase to 81%, compared to the overall average of 74%.
Sinch’s interpretation is that mature monitoring capabilities allow these teams to identify errors that less sophisticated organizations would simply miss.
“The most advanced organizations are not failing less, but failing faster,” Morris said. “A higher rollback rate improves monitoring and control rather than compromising performance.”
Investment in governance alone does not solve the problem
Data shows that companies are not ignoring this problem.
Investments in trust, security and compliance (76%) now reportedly outpace spending on AI development itself (63%), making it the single largest investment category in enterprise AI programs.
Here, Sinch introduces the concept of a “guardrail tax.” The idea is that safety infrastructure is becoming more critical and engineering capacity is increasingly being consumed. 84% of AI engineering teams reportedly spend at least half their time on safety systems instead of building new features or improving customer experiences.
For organizations under pressure to demonstrate AI ROI, this is a compounding cost with no clear endpoint.
Sinch’s data confirms that satisfaction with communications infrastructure is the strongest predictor of successful AI deployments, more so than governance maturity or overall level of investment. These conclusions align conveniently with Sinch’s own product offerings.
More than half of companies (55%) say they are building custom infrastructure simply to manage cross-channel context, and 86% are evaluating or actively considering switching communications providers.
Stakes keep rising
Despite the scale of the rollback and the resulting engineering burden, the appetite for AI investment shows no signs of slowing. 98% of companies report that they will increase AI communications spending in 2026. This means that the gap between ambition and sound execution will widen before it closes.
“Engineering teams spend most of their time building and maintaining safety systems, much of which must be delivered by communications infrastructure,” Morris added. “That’s a guardrail tax that slows down the organization.”
AI Production Paradox The early access report is available now, with full regional and industry analysis expected before the end of June.