Talking about AI in dollars

AI programs lose their funding when they cannot show what the work returned in numbers finance already tracks. Accuracy, latency, and F1 scores do not appear in any management report. The budget gets decided on revenue, gross margin, cycle time, working capital, headcount, and cost of capital. This is how we build the case and report against it in those terms.

The objective of the program

The sponsor states what the program is for in one sentence, before any use case is scored. Cutting cost to protect gross margin, growing revenue and share, and reaching technical parity with competitors who moved earlier are three different programs. They call for different use cases, different time horizons, and different tolerance for work that does not pay off.

A cost objective favors short use cases with savings a controller can trace to a cost center. A revenue objective supports longer and less certain work, and needs a board willing to wait for it. Portfolios assembled without that decision tend to contain some of each and defend none of them at the point where the budget is questioned.

Four sources of value

Use cases earn their funding on four things, and most business cases count only the first.

Operating cost. Hours removed from a task, shorter cycle time, lower external spend, less overtime, and in some cases volume absorbed without adding headcount.

Revenue and margin. Forecast accuracy that reduces stockouts and the working capital tied up in safety stock. Quoting and pricing that hold gross margin on deals that were being discounted out of habit. Service work that keeps revenue which was leaving.

Risk and compliance. Penalties avoided, exposure reduced, and the cost of an incident that a control prevented.

Reuse. The data work, integrations, and platform put in place for the first use case carry over, so each later one costs less and takes less time to deliver.

The fourth is where standard software ROI models get it wrong. They charge the whole foundation to the first project and credit none of the benefit to the ones that follow, which makes the data work look more expensive than it is and every early use case look marginal.

A funding decision made on the first use case alone will underfund the data and integration work every later use case depends on.

Measuring cost and return

Generative AI changed the unit of cost. Spend is no longer server time that stays roughly flat month to month. It is token consumption, which moves with prompt volume, context length, model choice, and how many steps an agent takes before it stops. Standard cloud cost management finds idle instances and sees none of this.

So normalize the cost to something the business recognizes. Instead of cost per token, report the cost to produce a thousand contract summaries or to process five thousand loan applications. Tag every call to a use case, a team, and a business unit so the spend has an owner. Then judge each use case on its return against the whole cost in the loop, the compute and the people both. Without that tagging, an agent that loops more than expected raises the bill with nobody responsible for noticing.

Funding early work

Net present value applied to a pilot usually kills it. The cost is immediate and certain, the revenue is later and uncertain, and the discount rate does the rest. Treat the early spend as buying the right to scale rather than the obligation to: the downside is capped at the cost of the pilot, and the decision to fund the full build is made after there is a measured result. That framing is familiar to any CFO who has approved a staged capital project.

Reporting to the board

Judge the portfolio rather than each project on its own, and put it in front of the CFO monthly: cost per business outcome, value captured with finance's validation, spend against forecast, and the use cases that were stopped and why. A monthly review lets the portfolio be corrected while the spend is still small.