Insights — July 2026
The unit economics of AI agents.
Agents change the cost curve of operational work — if you count all the costs. A framework for the arithmetic.
The business case for AI agents is not “headcount savings.” It is a change in the shape of the cost curve: work that used to cost more as volume grew starts costing roughly the same. Getting that case right means counting honestly, on both sides of the ledger.
The unit that matters
Pick the unit the business already understands: a processed invoice, a resolved ticket, a booked shipment. The agent’s job is to lower the cost of that unit and keep it flat as volume grows. If the workflow has no countable unit, it is a poor first candidate — start where you can measure.
What the unit cost includes
Model spend per task, amortized evaluation and monitoring, the human review share, and integration upkeep. A common mistake is to count only tokens. Tokens are usually the smallest line — the human exception rate is where the economics are won or lost. An agent that handles 60% of cases unattended, at a tenth of the manual cost, transforms the curve. One that needs review on every task just added a step.
The fixed costs are real
Evaluation harnesses, guardrails, integration, and data cleanup are the price of admission. They amortize over volume — which is why high-volume workflows pay back and low-volume ones do not. The break-even is arithmetic, not faith: fixed costs divided by per-unit savings equals the volume you need.
The curve, not the point
A single month’s savings is a data point. The case for agents is the trajectory: exception rates fall as the system learns your edge cases, model costs fall with each generation, and the same infrastructure serves the second workflow cheaper than the first. Price the program, not the pilot.