A pattern I see in budget discussions this year: model prices per token keep falling, yet the total AI bill keeps growing, and nobody in the room can say whether that is good news or bad news. It is the wrong argument, fought with the wrong number.
This is no longer a contrarian view. In a McKinsey Quarterly interview published this July, Pay-i’s CEO puts it flatly: per-token pricing has stopped being a useful measure of what enterprises actually pay for generative AI: consumption patterns, retries, orchestration overhead and agentic chains drive the real bill[C]. And the bill matters more than ever, because MIT’s State of AI in Business research finds 95 percent of corporate GenAI pilots deliver no measurable P&L impact: spending without a value counterpart[M].
The number that matters
Cost per token tells you what the vendor charges. It tells you nothing about what you get. The number that matters is cost per usable outcome: what it costs, end to end, to produce one result your business actually uses. A drafted contract clause that a lawyer accepts. A support answer a customer accepts. A forecast a decision is actually based on.
Measured that way, the maths often changes direction. A cheap model that produces drafts nobody trusts is expensive. A pricier model, or a human-in-the-loop process around a modest one, that produces accepted output is cheap. Agentic setups multiply this: an agent that loops, retries and calls tools can consume a hundred times the tokens of a simple prompt. If the outcome is real work completed, that can be excellent value. If nobody owns the outcome, you are buying very fast noise.
A worked example (illustrative)
Numbers below are deliberately simplified to show the logic, not to quote market prices.
| Cheap model, no review loop | Stronger model plus human check | |
|---|---|---|
| Cost per draft | 0.02 | 0.40 |
| Drafts accepted and used | 1 in 10 | 8 in 10 |
| Cost per usable outcome | 0.20 plus rework time | 0.50, no rework |
| Where the real cost sits | Human hours spent fixing output | Visible on the invoice, controllable |
What this means for you
Before you approve the next AI budget, ask for three numbers per use case: cost per usable outcome, acceptance rate of the output, and who owns the stop-decision if the numbers do not improve. If your team cannot produce those numbers yet, that is not a reason for embarrassment. It is the actual work, and it is worth doing before the bill grows another quarter.
The money is real. The measurement is not.
The scale of spending makes the measurement question urgent. Stanford’s 2026 AI Index counts 581.7 billion dollars of global corporate AI investment: up 130 percent in a single year, with generative AI investment alone growing 404 percent to 170.9 billion[ST]. Deloitte’s survey of 3,235 leaders adds the uncomfortable counterpart: spending keeps rising while ROI stays elusive for most, and only about a third of companies say AI is deeply transforming how they do business[DE]. Money at this scale deserves a real unit of account, and cost per token is not it.
What I see in the field
When I ask leadership teams what their AI actually costs per completed process, not per token, per outcome, the room usually goes quiet. Mid-sized companies feel this first because nobody there has slack budget to hide it. The fix is not a cheaper model. It is cost modelling at the level where value is created: the workflow.
Honest AI cost modelling is a standing part of our work: cost per outcome, not cost per token, mapped against where AI genuinely moves your results. If your AI bill rises while your prices per token fall, we should talk.
Personalised answer within 48 hours.
Sources
- McKinsey Quarterly, Cost versus Value: Managing Agentic AI System Performance (interview with David Tepper, Pay-i), July 2026. mckinsey.com
- MIT NANDA, The GenAI Divide: State of AI in Business 2025; reporting by Fortune, August 2025. fortune.com
- Stanford HAI, The 2026 AI Index Report. hai.stanford.edu
- Deloitte, The State of AI in the Enterprise 2026 (survey of 3,235 leaders). deloitte.com
The AI Act Just Moved. Your Deadline Did Not.