The Reversal Rate

Sinch published a number this week that should end a few conversations. Seventy-four percent of enterprise AI agent deployments have been reversed after going live. Not during pilot. After production. The agents worked. The organizations did not know what to do once they were running.

I wrote about this today because I see the same failure pattern from the inside. Teams pick a model, build the agent, ship it. Then nobody owns what happens next. No escalation path. No authority boundary. No one watching for silent failures. The governance layer that tells the agent where to stop does not exist. So the whole thing comes back.

The 26% that kept their deployments running built that layer before they needed it. Monitoring on day one. Explicit ownership. Defined edges. Not better models — better operations.

Meanwhile, a UJET study found zero contact center agents describe AI as essential. Zero. These people use AI tools every shift. The technology is present. It has not changed the work underneath it. That gap is not a model problem. It is a deployment design problem.

The industry spent two years asking “can we deploy this.” The question that actually mattered was “can we govern this once it is live.” Seventy-four percent found out the hard way.