Sinch: 74% of firms pulled agents

AI agents for customer service were supposed to be the easy win. They answer chats, route messages, and cut support costs. But the new picture is rougher than the hype. Sinch said on May 13 that 74% of enterprises have already rolled back or shut down a live AI customer-communications agent after deployment, and Akeyless said a day earlier that two-thirds of organizations suspect AI agents have already accessed data they were never meant to touch. ### What exactly did Sinch say? Sinch’s new report, *The AI Production Paradox*, is about customer communications specifically — the bots and agents that talk to customers over chat, messaging, and similar channels. The big surprise is not that companies are experimenting. It’s that many are already past the pilot stage. Sinch said 62% have AI customer-communications agents live in production, but 74% have also rolled one back or shut one down after deployment because of a governance failure. (prnewswire.com) The survey covered 2,527 senior decision-makers across 10 countries and six industries. ### Why is that number so weird? Because the firms with the best guardrails reported even more pullbacks, not fewer. Sinch said the rollback rate rises to 81% among organizations with fully mature guardrails. That sounds backward until you think about what “mature guardrails” really means. It often means a company is better at noticing when an agent is going off-policy, mishandling an interaction, or creating compliance risk — so the company actually pulls the plug instead of letting the problem linger. (prnewswire.com) That’s less a failure of caution than proof that monitoring is catching things. ### What counts as a governance failure? Basically, not the model being dumb in a vacuum. It’s the surrounding system failing. The agent says something it should not say, takes an action it should not take, touches data it should not see, or operates without a clear approval and escalation path. In customer communications, that gets dangerous fast — one bad reply can become a compliance issue, a privacy issue, or a brand issue in front of thousands of users. (prnewswire.com) Sinch framed the problem as systemic failures after deployment, which is the key point here. ### Where does Akeyless fit in? Akeyless looked at the identity and access side — the part that decides what an AI agent is allowed to reach. Its May 12 study said 67% of organizations using AI agents suspect those agents have already accessed data beyond intended scope. The same release said many companies cannot detect compromised agents for hours and are already spending more than $1 million dealing with the fallout. The picture is simple: agents move on machine timescales, but many defenses still run on human review timescales. (prnewswire.com) ### Why does identity matter so much? Because an AI agent with the wrong credentials is like an intern who somehow got the master keys. The model does not need to “hack” anything if the system hands it broad access by default. Once an agent can read customer records, internal docs, or transaction systems, the main risk is not hallucination. It’s authorized misuse — accidental or otherwise. Akeyless is pushing that exact point with its framing around AI agent identity security. (akeyless.io) ### So is this about bad models? Not mainly. The pattern across both reports points somewhere else. Companies are getting agents into production. Then the real trouble starts — permissions, oversight, escalation, audit trails, and policy boundaries. In other words, the hard part is not making the bot speak. The hard part is making the whole system behave. (akeyless.io) ### What should readers take from this? The AI-agent story has moved past “can it work?” and into “can you control it after launch?” That is a much tougher question. The firms pulling agents back are not necessarily losing the race. They may just be the ones discovering, earlier than others, that production AI fails at the edges — where trust, access, and accountability live. (prnewswire.com)