The "Human Throttle" Problem That's Killing Enterprise AI Agent ROI

Nate B. Jones argues that the real bottleneck preventing enterprise AI agents from taking on meaningful autonomous work is not model intelligence — it’s trust, and trust is fundamentally a function of reversibility. Drawing on Amazon’s ‘two-way door’ and ‘one-way door’ framework, he explains why AI agents have made the fastest inroads in software engineering specifically: decades of investment in Git, code review, automated testing, staged rollouts, and monitoring have made software mistakes survivable and correctable at machine speed. That safety infrastructure is largely absent from the rest of the business.

In domains like finance, HR, legal, compliance, and customer communications, many consequential actions cannot be easily undone. Humans naturally introduce friction — hesitation, double-checking, social anxiety about errors — that acts as an informal safety system. When agents operate at machine speed with no reputational risk, that informal brake disappears, and the consequences of errors can become irreversible before anyone notices.

Jones lays out a practical set of primitives borrowed from software culture that business leaders can adopt to expand agent autonomy safely: draft-first (no important action goes straight from idea to done — it enters a proposed state first), preview as a primitive (every agent action can be inspected before it finalizes), and staged rollouts applied to non-software decisions. His conclusion is that the most important enterprise AI work in 2026 is not selecting the right model — it is redesigning business processes so that more decisions become safely reversible, which is the prerequisite for unlocking real agent ROI.

📺 Source: AI News & Strategy Daily | Nate B Jones · Published December 29, 2025
🏷️ Format: Deep Dive