The Settlement Engine

Every company decides once a year what its workforce is worth to it: which capabilities to build to an edge, which to hold at competent-average, and where to place the disproportionate bet. Those bets are placed against half a ledger. Workforce cost is reconciled to the dollar. What the workforce produced, net of what a competent standard performer would have produced under the same conditions, is not kept at all. Agentic AI does not change those bets. It settles them. Where a company invested to win, AI compounds the capability. Where it chose average on purpose, AI commoditizes the function and the saving is real. Where it invested in nothing it could name, AI erodes the work that carried the margin. Two of the three are wins and the third is a loss, and on a ledger that records cost and not value all three arrive as the same thing: a cost line that is steady or falling. This paper traces one familiar failure to the distinction beneath it. The multi-million-dollar compensation play, built on the only number a company can defend, moves the attrition figure and still cannot prove it paid for itself, because nothing the company keeps separates the raise from the demand shock that arrived alongside it. The comp survey prices the role against the market. The value at risk is the work. Those are different axes, and the spend is aimed at the wrong one. The direction of AI's distributional effect is treated here as a design variable rather than a property of AI: set by whether the work has a checkable right answer, whether the context that makes it good has been captured in a form the tool can be given, whether judging the output takes the expertise being augmented, and whether the agent suggests to a person or works as a coworker. Where the model is reliably competent, the floor rises and the distribution compresses. Where the work sits at or past the edge of what the model does well, the gain goes to whoever can tell a right answer from a plausible one. The published evidence runs both ways because the conditions do. The requirement is a single addition: the other half of the ledger, governed. It is a capital allocation method rather than a pay-setting method. The paper closes on where that number is settled whether a company priced it or not, which is the exit.