Why AI ROI keeps disappearing between the desk and the P&L ?

Why the same company can show employees working four to eight times faster while finance can’t find the return anywhere on the income statement.

The specific mechanical reason, not “governance,” not “change management,” that explains why AI ROI keeps disappearing between the desk and the P&L.

Why Anthropic’s own account of its internal engineering operations is the cleanest real world proof of this dynamic available today.

A reusable diagnostic question for judging whether your next AI purchase will move your bottom line or just relocate a bottleneck.

Which sectors and vendors are structurally positioned to capture AI ROI first, and why the rest are stuck.

Why quietly cutting entry level roles to “save on AI-replaceable work” often makes this exact problem worse.

Individual AI productivity gains are real and large: engineers merging several times more code, knowledge workers reclaiming hours per week.

Organizational ROI hasn’t kept pace. Multiple 2026 surveys put “significant” ROI from generative AI or agents in the 20s percent, far below the share reporting some benefit.

Weak governance and immature strategy are real but secondary causes. The primary one is structural: AI speeds up one stage of a workflow and leaves the rest exactly as slow as before.

The untouched stage doesn’t just cap the gain, it absorbs a flood of new AI-generated volume and gets measurably more expensive to run.

The fix isn’t more AI adoption. It’s redesigning whichever stage AI didn’t touch, usually verification, judgment, and sign-off.