Nearly three in four enterprises that deployed an AI agent to communicate with customers have been forced to roll it back or shut it down, according to new research from Sinch, and the organizations with the most sophisticated governance programs are failing at an even higher rate.
The finding matters for an accounts receivable management industry that has spent the past two years standing up chatbots, virtual agents, and automated outreach tools while navigating Reg F and state-level compliance requirements. The message from the data: getting an AI agent live is no longer the hard part. Keeping it live is.
The AI Production Paradox report, based on a January 2026 survey of 2,527 director-level and above decision makers across 10 countries and six industries, found that 62% of enterprises already have AI agents in production across customer channels, with 88% expecting to be there within 12 months. But 74% of those that shipped have had to pull an agent back due to a governance failure. Among organizations describing their guardrails as fully mature, the rollback rate climbs to 81%.
The leading causes should sound familiar to anyone in collections compliance. Customer data exposure was cited by 31% of organizations that experienced a rollback, followed by hallucinations or brand risk at 22%. Another 16% could not fully diagnose what went wrong because no audit trail existed, leaving them unable to prove the problem was fixed.
In financial services specifically, the rollback rate sits at 69%, and reputational damage outranks support queue overload as the top fear when an agent fails. Fraud prevention ranked as the industry’s second-most important AI deployment goal at 43%, higher than any other sector surveyed.
Sinch’s Chief Product Officer Daniel Morris argues the higher failure rate among mature programs reflects better monitoring, not worse performance. Organizations with strong instrumentation catch failures that others simply never see. The report suggests the companies reporting zero rollbacks may be the ones with the least visibility into what their agents are actually doing.
The research points to infrastructure, not governance spending, as the differentiator. Trust, security, and compliance is the top investment category globally at 75%, ahead of AI development itself at 63%, yet failure rates have not budged. Meanwhile, 84% of engineering teams report spending at least half their time building guardrails rather than improving functionality, and 55% of enterprises are custom-building the ability to preserve customer context when a consumer switches channels. For agencies fielding consumer contacts across text, email, voice, and chat, that context gap is where a technical failure becomes a compliance failure.
The market is responding: 86% of enterprises have had active or exploratory conversations with alternative communications providers in the past 12 months.
