Who Actually Gets Paid in the Agentic Economy

There is a version of the agentic AI story where enterprises are the clear winners: they deploy autonomous agents, cut costs, and capture the productivity gains. There is a less comfortable version underneath it, where the consumption based economics of most agent deployments function less like a productivity purchase and more like a rental arrangement that funnels an outsized share of the value upstream, toward compute and model providers, rather than toward the enterprise doing the deploying.

Governance is the first place this shows up. Agents make spending decisions autonomously, often faster than the procurement processes built to review them were ever designed to move. It is entirely possible for a pilot deployment to triple its monthly consumption inside a normal reporting cycle without anyone signing off on that specific increase, because nobody built a control that would have caught it in time.

The cost curve compounds the problem. Token consumption doesn't scale linearly with the value an agent delivers: reasoning heavy models and multi agent orchestration in particular can produce cost growth that outpaces the productivity gain finance teams modeled going in, and most budget forecasts still assume something closer to a flat, linear relationship.

The structural risk sits underneath both of those: enterprises using agents specifically to reduce headcount are, in aggregate, shrinking the very customer base that funded the software industry's growth for two decades, while a narrow set of compute and model providers capture a growing share of the value being created. The available correction is to move from consumption denominated to outcome denominated contracts, put board level controls on autonomous spend rather than department level ones, and model the demand side consequences of an agent deployment, not just the cost line it's supposed to shrink.