Talking about AI in dollars

An AI program lives or dies on whether finance believes its numbers. Engineers report accuracy, latency, and F1 scores. The CFO reports return, risk, and cost of capital. When the two never meet, the program looks busy, the board grows skeptical, and the funding quietly dries up. Here is how to make the case in the language the board decides in.

Name the corporate wager first

Before a single use case is scored, the sponsor states the overarching bet in one sentence. There are three honest options, and they are not the same program. Defensive — protect margin through cost reduction. Offensive — capture share and grow revenue. Parity — reach technological par with disruptors already moving.

The choice sets the prioritization rule, the time horizon, and the risk appetite. A defensive bet favors quick wins with hard savings; an offensive bet tolerates longer, riskier lighthouse work. Say it out loud, because a portfolio built for the wrong wager will look active and still disappoint the board.

Justify value across four layers, not one

Every use case earns its funding on four layers. Operational efficiency — the hard savings from faster completion and lower external spend. Revenue lift — forecasting that prevents stockouts, pricing that captures margin, engagement that reduces churn. Risk and compliance — avoided penalties, mitigated exposure, brand protected from algorithmic failure.

The fourth is the one traditional software ROI misses, and over time it is the most important. Call it the architectural dividend. When you build a unified data foundation for the first use case, the next one reuses it — so its marginal cost drops and its time to value shrinks. The first application pays full freight; the tenth rides on infrastructure already built.

A board that funds only on first-use-case ROI will underfund the foundation and never see the compounding. Show the dividend explicitly.

Report return on AI, not cost per token

Generative AI breaks the old unit of measurement. Cloud cost is no longer stable server time — it is token consumption, which swings with prompt volume, context length, model choice, and how many autonomous loops an agent runs. Standard cloud cost management catches idle servers and sees none of this.

So normalize cost to a business outcome. Don't report cost per token; report the cost to generate a thousand contract summaries or process five thousand loan applications. Tag every call to a use case, a pod, and a business unit, then judge each on return on AI — economic return over the sum of the human and the compute cost in the loop. It keeps compute intensity tied to commercial value, and stops a runaway agent from quietly eating the budget.

Fund the portfolio like options, and report it monthly

Net present value kills strategic bets too early, because the upfront cost is high and near-term revenue is uncertain. Treat early spend differently: funding a pilot or a data layer buys the right, not the obligation, to scale later once the value is proven. Price it as an option — you keep the upside and cap the downside at the cost of the pilot.

Then judge the portfolio, not each project alone, on a live dashboard the CFO can read: cost per business outcome, return on AI, finance-validated value captured, and token spend against forecast. The single strongest predictor of returns is whether AI is a standing board agenda item rather than an annual review. A board that engages monthly is the one that sees returns. A board that reviews AI once a year is the one that writes off its investment.