Accenture named the reason cheaper AI produces bigger bills.
Give a powerful model to a whole workforce and it thinks more, branches more, and calls more tools. The efficiency becomes the cost.
The most useful idea any consulting firm shipped this week came with a name attached. Accenture calls it the agentic Jevons paradox, after the nineteenth-century observation that making a resource cheaper to use tends to increase, not decrease, how much of it gets consumed. Applied to AI the mechanism is exact. Hand a powerful model to an entire workforce and people default to the most capable and most expensive option, ask it more often, and let it call more tools and spawn more branches to answer. Each of those choices is individually reasonable. Multiplied across a company they are where the spend quietly balloons, which is why a falling price per token keeps arriving as a rising invoice.
Lan Guan, Accenture's chief AI and data officer, paired the diagnosis with a discipline: see the usage, treat it, then manage it, and measure success not by cost per token but by value delivered. It is a clean framework and an honest one, and it is also, almost word for word, the argument this publication has been making for a year. When the largest consulting firm on earth tells CFOs that the point is value delivered rather than cost per token, the debate about whether AI needs an owned measurement layer is effectively over. The only open question is who supplies it.
That is where the Token Navigator becomes a tell. It is a practice that instruments your usage, tunes which model each request is routed to, and reports back on a bill that still belongs to the supplier and a discipline that still belongs to Accenture. The savings are real. They are also a subscription to the firm that reads your meter, and the moment the underlying model prices move, and they move constantly now, the tuning has to be redone and rebilled. You have not removed the dependency. You have added a second one on top of it, and called the second one governance.
The alternative is not to read the meter better. It is to take most of the work off the meter. The deterministic core of a workflow, the extraction, the validation, the reconciliation, the routing rules, does not require a model once it is encoded as software the client owns, and code that does not call a model does not consume a token no matter how the price moves. Reserve the model for the genuine judgment call, meter only that, and the Jevons paradox loses its surface area, because the workforce can no longer spend without limit on the ninety percent of the task that was never a judgment in the first place.
Accenture has done the field a service by naming the trap in public. The distance between its answer and ours is the whole business. Accenture sells you the instrument to watch the cost. Consulting-as-Code removes the cost from the deterministic work entirely, and hands you the record instead of the report.