I sit in a lot of conversations about AI return on investment. Most of them go badly. The reason is usually that the conversation has been framed in a unit that does not match the actual economics of the system being discussed.
The dominant frame is wrong for most systems
The default frame for AI ROI is total project cost against projected savings, evaluated annually, on a discounted cash flow basis. This is how a CFO is trained to evaluate any project. It is not the right frame for most AI deployments.
The reason is that AI deployments tend to have unit economics that change rapidly over time. Inference costs fall. Capabilities expand. The same workload that costs ten cents per task today might cost two cents per task next year, and the model that runs it might be capable of an additional task adjacent to it.
A project-level ROI evaluation collapses this dynamic. It treats the system as a static investment with a fixed cost profile. The result is that good systems look bad on paper, because the depreciation schedule does not capture the trajectory of the underlying technology.
The frame that works
The frame I have moved my clients toward is unit economics per task, measured monthly, with a deliberate trajectory projection.
For each task the system performs, calculate the all-in cost of having the system perform it, including inference, infra, labour for monitoring, and amortised development. Compare against the all-in cost of the manual baseline. The ratio is the unit economic.
Project the trajectory of that ratio over twelve to twenty-four months. Inference costs are likely to fall. Manual costs are unlikely to fall. The unit economic is likely to improve. The project decision is whether the current ratio justifies the deployment, with confidence that the future ratio will be better.
This frame produces decisions that do not collapse when the technology underneath shifts. It also forces a discipline that the project-level frame does not. You have to define the task unit. You have to measure the manual baseline. You have to build the monitoring that produces the unit economic in production.
The questions that fall out of this frame
A few questions that this frame surfaces and the project-level frame buries.
What is the marginal cost of the next task. If your unit economic improves with volume, the deployment has operational leverage. If it does not, the deployment is essentially a hosted utility, and the value capture is different.
What is the cost of a wrong task, weighted by the rate of wrongness. A system that is cheap per task but produces costly errors at a meaningful rate has a unit economic that is much worse than the headline number suggests. The error cost has to be in the calculation.
What is the human review overhead, expressed as a tax on the unit economic. Most AI systems in regulated domains have human review somewhere. The labour cost of that review is real. It is sometimes more than the inference cost. It belongs in the unit economic.
What is the trajectory of the manual baseline. If the manual baseline is also improving, through process redesign or training, the AI unit economic has to improve faster to win. This is sometimes overlooked.
What this frame does to the conversation
The conversation gets more honest. The deployments that have genuine unit economics shine. The deployments that were financed on enthusiasm reveal themselves. The conversation about funding shifts from "is this project worth it" to "is the unit economic improving and, if not, what is the constraint".
That second question is much more useful. It produces specific engineering work. It is testable. It has a path to resolution that the first question does not.
A note on time horizons
A nuance worth naming. Some AI deployments have unit economics that only work at scale. The cost of building the system is high, the marginal cost per task is low, and the break-even point is at high volume.
For these deployments, the unit-economic frame still applies, but the project decision is partly a bet on volume. The question becomes whether you can get to break-even volume, and how confident you are. This is a different conversation than the one about steady-state economics, and the two should not be conflated.
The CFOs I work with are usually willing to fund either type of deployment. They are not willing to fund a deployment that cannot tell them which type it is.