July 22, 2026
Most Enterprises Have Had to Pull a Live AI Agent Offline, Report Reveals
For the past two years, AI agents have been sold as the future of customer communications. Vendors praised them, budgets grew around them, and by now most enterprises have moved past the pilot stage entirely. On paper, the technology has arrived.
According to a new report from Sinch, 62% of organisations already have AI agents live across customer channels, with 88% expecting to be in production by the end of 2026. Yet 74% of those that shipped an agent have since been forced to roll it back or shut it down. The failure rate holds across every region and every industry surveyed, from 66% in technology to 85% in professional services. The question the industry now has to answer is not whether AI agents work. It is why so many of them are being pulled out of production.
What is Going Wrong
The leading cause is personal data exposure, cited by 31% of organisations that reported a rollback. Hallucinations and brand risk come second, meaning an agent gave a customer confidently wrong information on a live channel.
These failures happen in front of customers, in real interactions, under the company’s own name, so the consequences go beyond a bad transcript. Regulators in the US, EU, and Canada have all confirmed that companies remain liable for what their chatbots tell customers, meaning a failed interaction is no longer just an embarrassment, but a legal problem.
What makes the rollback rate harder to explain is that agents are not being asked to do very much in the first place. Infobip’s recent report shows that half of enterprises globally have deployed agentic AI, yet that adoption is wide but shallow. Most of those deployments handle simple triggers such as reminders, notifications, and identity checks, while the multi-step journeys that justify autonomous reasoning stay in manual hands. Product returns and refunds, among the most involved journeys of all, is automated at just 15%. Agents are being pointed at the low-stakes work, and the majority of the organisations that shipped one have still had to take it down.
More Governance, More Rollbacks
The most concerning finding is that the organisations doing everything “right” fail more often, not less. Among companies that describe their guardrails as fully mature, the rollback rate rises to 81%, according to the Sinch data. The report’s explanation is that these teams are not running worse programmes. They simply have the monitoring in place to catch failures that less instrumented organisations never see. By that logic, the companies reporting zero rollbacks may be the ones with the least visibility into their own systems.
Confidence offers no protection either, as 90% of decision-makers describe themselves as confident in their AI readiness, yet 75% of that confident group has experienced at least one governance rollback.
When an agent goes down, 35% of organisations say the biggest impact is a surge in human support load, as every conversation the AI was handling reverts to people. Another third cite reputational damage and loss of customer trust, which has no quick fix. The third cost is engineering time: the majority of AI engineering teams report spending at least half their time building guardrails and safety controls rather than improving the agent itself.
That last figure should give pause to any team still treating deployment as the finish line. Most chatbot problems trace back to the design and maintenance work around the bot, not the platform itself, and the rollback data suggests the same lesson applies to agentic AI. Getting an agent live was supposed to be the hard part. Keeping it live, it turns out, is where the real work begins.
