Assurance Has a Half-Life: Why Continual Learning Breaks Point-in-Time Governance for Enterprise AI Agents

The instruments we trust to certify enterprise AI agents—audits, red-team exercises, evaluation benchmarks, certifications, and regulatory conformity assessments—share one hidden assumption: that the system assured today is the system running tomorrow. Continual learning voids it. An agent built to adapt to new data, refine its tool use, and update its memory after deployment is, by construction, no longer the artifact that was assured, and the gap may widen as relevant adaptations accumulate. I argue that assurance evidence therefore has a half-life: an interval over which it faithfully describes the running system, finite for an adaptive agent and—the crux—almost never declared. The object that decays is the governance artifact, not the model; this is what separates the problem from the familiar one of model drift. I make the argument precise, give a compact anatomy of what erodes an assurance claim, ground it in the public record of real AI incidents, show that both technical and regulatory governance silently inherit the point-in-time assumption, and close with a research agenda for treating assurance as perishable. This is a position paper: I propose no re-assurance mechanism, and I argue that stating the problem correctly is the prerequisite the enterprise-agent setting currently lacks.