A new global study from communications platform Sinch puts a number on something a lot of businesses are experiencing but few are talking about openly: 74% of enterprises have already rolled back or shut down a live AI customer communications agent after deployment.

The report, titled “The AI Production Paradox,” surveyed 2,527 senior decision-makers across 10 countries and six industries. The picture it paints is not what the industry has been selling.

The Problem Has Shifted

For years, the conversation around enterprise AI has centred on adoption: how do you get a pilot off the ground, how do you convince leadership, how do you clear IT and compliance hurdles to reach production? The Sinch data suggests that conversation is now out of date.

According to the research, 62% of enterprises already have AI agents running live in production. Deployment is no longer the hard part. Keeping them running well is.

The study found that failures in governance, performance consistency, and runtime control are what’s driving rollbacks. Businesses discover, often after the fact, that their AI agents behave differently in the wild than they did in testing. Edge cases appear. Hallucinations surface in customer-facing contexts. Agents escalate things they should handle, or handle things they should escalate.

The Paradox in the Data

One finding in particular stands out: rollback rates are actually highest among the organisations with the most mature governance frameworks. While the overall rollback rate is 74%, it climbs to 81% among enterprises with fully mature guardrails and monitoring in place.

That sounds counterintuitive at first. But Sinch CPO Daniel Morris offers a reasonable explanation: “Higher rollback rates reflect better monitoring and control, not weaker performance.”

In other words, organisations that invest in visibility catch problems that other organisations simply miss. Less mature businesses may have agents quietly underperforming without anyone noticing, while more sophisticated operations pull the plug the moment something drifts.

This is not a story about AI failing. It is a story about what it takes to run AI responsibly at scale.

Infrastructure as the Real Differentiator

The study identifies communications infrastructure satisfaction as the strongest predictor of successful AI deployment outcomes, more so than investment levels or governance maturity. Eighty-seven percent of organisations rate high-performance infrastructure as essential or very important.

This matters for businesses thinking about AI agent deployments. The model is only one piece of the puzzle. The underlying infrastructure, the data pipelines, the integration architecture, the monitoring tooling, these determine whether an agent holds up under real-world load.

Investment is Not Slowing Down

Here is the other half of the paradox: despite 74% rollback rates, 98% of enterprises report increasing their investment in AI communications in 2026.

The industry is not retreating. If anything, the rollbacks are functioning as a forcing mechanism. Organisations are learning faster, resetting expectations, and coming back with more realistic architectures.

For businesses that have not yet deployed, the lesson is clear: the cost of a poor deployment is not just the rollback. It is the trust deficit with your team, your customers, and your leadership. Getting the foundations right before you go live matters more than getting there first.

What This Means for Business

The Sinch research documents a pattern that experienced practitioners have seen firsthand: production is where AI gets real, and real is harder than demo.

For businesses planning their first AI agent deployment, the research points to a few things worth internalising:

Governance is not optional, but it is not enough on its own. Having guardrails means you will catch problems. It does not mean the problems will not happen. Plan for rollback scenarios from day one.

Infrastructure is a competitive advantage. The organisations with the best outcomes are not necessarily the ones with the best models. They are the ones with the best plumbing.

Rollbacks are not failures. Pulling an agent that is not performing is a sign of operational maturity, not weakness. The organisations in this study that have rolled back and redeployed are almost certainly building something more durable the second time.

If your business is deploying AI agents to handle customer communications, field enquiries, manage bookings, or run internal workflows, the production phase is where your real design choices get tested. The businesses that navigate it well do so with a clear architecture, proper monitoring, and a realistic understanding of what a live AI system actually demands.

Enterprise DNA works with businesses at exactly this stage. Whether you need help designing an agent architecture that holds up in production, or an outside perspective on where your current deployment may be drifting, that is the kind of work our Omni Advisory service is built for.