A new JumpCloud survey, reported by BankInfoSecurity, puts a number on a problem most coverage has treated abstractly. More than half of organizations — 55% — now use or test AI agents capable of changing systems, permissions, records, or workflows. But among those already running agents, 59% haven't fully extended their human identity and access management policies to cover them. That gap matters more than it sounds like it should. Forrester's Andras Cser frames the issue precisely: AI agents behave like "smarter, less deterministic APIs," and the combination of scale and autonomy creates an amplifying effect at the enterprise level. Unlike a service account calling a fixed set of endpoints, an agent can interpret a goal, choose its own tools, and take actions nobody explicitly programmed. Enterprises can end up running more agent identities than human ones, each with more latitude to act independently than a traditional machine account ever had. The fix Cser proposes isn't a new framework bolted onto AI specifically — it's extending the IAM discipline that already exists. Treat agents inside existing identity and access controls rather than building a parallel governance track for them. That's a less glamorous answer than a new protocol or standard, but it's also the one most directly actionable for CIOs and CISOs today: give every agent a distinct identity, limit its access, and tie its actions back to an accountable human owner, using the same control plane already governing employees and service accounts. The survey data suggests most organizations know this in principle and haven't done it in practice. Deployment is outpacing governance by a wide margin — agents are live and making changes in production systems at over half of surveyed companies, while the majority of those same companies haven't closed the basic IAM extension that would make those agents accountable in the same way a human employee is. That's not a future risk; it's a current operational gap, and it's the kind of gap that tends to surface during an incident rather than an audit. What's notable here isn't a new technical spec — it's confirmation, via survey data rather than vendor pitch, that the identity foundation for agent governance is still unbuilt at most companies, even as adoption accelerates past it.