The enterprise AI problem was supposed to be getting out of pilot mode. Turns out, it’s what happens after you get out.
A global survey from communications platform Sinch found that 74% of enterprises have already rolled back or shut down a customer-facing AI agent after deployment. Among organizations with the most mature governance frameworks, the rate climbs to 81%.
The data, drawn from 2,527 senior decision-makers across 10 countries including Canada, shows that most enterprises have already moved AI agents into production.
The trouble starts after launch.
Customer data exposure was the leading trigger for rollbacks, followed by hallucination or brand risk and, in 16% of cases, the inability to diagnose what went wrong at all.
Organizations with the most sophisticated guardrails rolled back at higher rates than those with less mature frameworks.
“The most advanced organizations aren’t failing less; they’re seeing failures sooner,” said Daniel Morris, Sinch’s chief product officer. “Higher rollback rates reflect better monitoring and control, not weaker performance.”
The organizations with weaker monitoring may be having the same problems and not catching them.
Greg Carlucci, senior director analyst at Gartner, told CX Dive that AI agents require unified, organized, and centralized data to function properly, and most organizations have not built that foundation yet. Pre-launch testing can only go so far, he said.
“Until you actually see it in use with the customer, there are things that you’ll need to adjust,” he said.
At a Toronto Tech Week panel covered by Digital Journal, AskCipher co-founder Shayan Rastgou said people either expect flawless performance or human-level reasoning from AI agents, and they don’t respond well when they get neither.
“They become angry, and then they get into their seats, and they yell at the AI, and the AI won’t do anything good for them,” Rastgou said.
Even so, enterprises are pushing forward. Sinch found 98% are increasing AI investment in 2026, and 62% already have agents in production despite the 74% rollback rate.
So where are the bottlenecks? Engineering teams are spending the majority of their time building and maintaining safety infrastructure rather than improving the product.
Morris called this the “guardrail tax.”
And communications infrastructure satisfaction was a stronger predictor of successful deployment than either investment levels or governance maturity, suggesting the problem lives below the governance layer, in the plumbing most deployment plans skip.
The organizations getting past the rollback cycle are the ones investing there first.
Final shots
- The rollback rate climbs to 81% among organizations with mature governance. They’re the ones with enough visibility to know when something breaks.
- Sinch found 62% of enterprises already have AI agents in production, even as 74% have rolled back or shut down a deployed agent. And 98% say they’re increasing AI investment in 2026 regardless.
- Infrastructure satisfaction predicts successful deployment better than guardrail maturity or spending levels. The organizations getting AI agents to stick are fixing what’s underneath.