Token costs used to be an engineering problem. They're now a board conversation. Every CFO we talk to is running the same internal exercise: how much are we spending on AI, per workflow, per team, per outcome — and how does that compare to the value we're claiming?
Most organisations can't answer this. The data exists in fragments — cloud bills, provider invoices, internal chargebacks — but joining it to a business outcome takes a week of work every time someone asks.
Why cost-per-task is the only metric that ends arguments
Total AI spend is a number. It doesn't tell you anything on its own. A $2M AI bill is either cheap or expensive depending on what it produced — that's the comparison that actually matters.
Cost per task — or per interaction, per resolution, per document processed — puts AI on the same footing as the rest of your operations. Your call centre has a cost per call. Your ops team has a cost per ticket. AI should have one too, and it should be auto-generated, not reconstructed each quarter.
What you need to attribute spend properly
Three things. A ledger that records every model call with the use case, team, and workflow it belongs to. A join to your budget structure so costs roll up to the right cost centres. And a gateway that enforces budget caps when they're reached — otherwise the ledger is just a recording mechanism, not a control.
If you have the ledger but not the caps, costs can run ahead of the conversation. If you have caps but no ledger, you can't defend the numbers. You need both.
Three questions your CFO should be able to answer
What does this workflow cost us per interaction this month, and how has that trended? What's the cost difference between our first and second use case, and why? If we double usage, what does the bill look like — linear, super-linear, or sub-linear?
If the answers take more than a day, your cost plane isn't instrumented. That's the gap to close, and it closes through the Cost Lens layer — not through more spreadsheets.
Unit economics beat ROI stories
ROI decks are the easiest thing in the world to construct and the hardest thing to trust. Unit economics — cost per interaction, margin per resolution — are harder to assemble but far more defensible. The first gets your initiative funded. The second gets your initiative renewed.
Build toward the second from day one, even if the numbers are rough. By the time the exec sponsor asks, the machinery to answer is already running.