Illustrative. Assumes quality remains within its agreed-upon tolerance; the unit cost matters only if the defect rate holds.

You were sold intelligence by the token. You should manage it by the outcome.
Same spend curve, two completely different stories, and the only difference is whether there's a number underneath. The good news: The raw material is already sitting in your billing data. Eighty to 90% of AI spend is inference now, and most of it can be tagged to a workflow.4 It's all just waiting for a definition.

Start this quarter

Four moves. The cheapest line item in your AI budget and the only one that explains the rest.

Give three workflows a denominator: Take your three biggest-spending agent workflows and name the outcome unit for each one. If you can't name what a workflow produces, that's a finding about the workflow, not the metric.
Measure cost per outcome, even if it's rough: The attribution won't be clean on day one. Split shared costs roughly and sharpen them every quarter. A decent approximation beats a perfect blank.
Compare it to the human-only cost: Fully loaded, including review time. This is the number that turns your AI report from an expense summary into a business case.
Put the trendline in front of your CFO: One page, three workflows, unit cost over time. Then let the metric make the calls, including which workflow to switch off. The invoice never could.

It took airlines a couple of decades to land on the metric that ended up steering the whole industry. The agentic era isn't going to wait that long. The advantage goes to whoever builds the denominator first, because companies that can't say what a decision costs will keep capping the machine, while the ones that can will keep feeding it.

Spend this quarter on the definition. It's the cheapest line item in your entire AI budget and the only one that explains the rest.

Appendix

Why not put value on top?

A fair pushback on this whole piece: Wouldn't it be more logical to flip the fraction value on top, spend underneath? Return on every AI dollar. That's the number a board really wants, after all. And yes, it's the right destination. It's just the wrong first metric. Three reasons.

One: you can't actually compute it yet. Value ÷ spend means putting a dollar figure on the value first and what's a resolved ticket worth? What did that credit memo earn? That valuation exercise is exactly what's kept AI ROI unmeasured for three years. A metric that waits on it never gets built. Cost per decision flips the difficulty: Both halves are countable today, the spend is in your billing data, and the outcomes are in your workflow systems.
Two: it's how everyone else does it. Every unit-economics tradition this article borrows from runs the same way: CASM is cost per seat mile, not seat miles per dollar. Manufacturing, cost per unit. Logistics, cost per delivery. The convention stuck because cost-per-unit is the version you can take apart when the number moves.
Three: one leads to the other anyway. The two ratios aren't rivals; cost per decision is the stepping stone. Once the outcome unit exists, value per decision is one multiplication away. And that human-only baseline in move 03 is quietly a value proxy: If a human decision costs 40 USD fully loaded and the agents cost 6 USD, the return story writes itself, no monetizing required. Build the denominator first; the number your board wants falls out of it.

Sources

Enterprise API Analysis. "Blended Enterprise Cost of AI (Q1 2025 – Q1 2026)."
FinOps Foundation. State of FinOps 2026.
Internal Reference: AI Masterclass Finance. Virtusa, 2026.
Stanford Human-Centered AI (HAI). AI Index Report. Stanford University, 2026.

Services

Industries

Insights

Careers

About