The cheap-token era is over, and consumption bills are outrunning the AI business case

The gap inside IBM's finding is the part worth taking to a budget meeting: 53 percentage points between the projected return and the actual one, on the same project ^[10]. No productivity claim being made for AI tooling is that large. A business case with an error bar that wide is not a forecast, it is a hope with a spreadsheet attached. And note which word is load-bearing in IBM's framing: unaccounted ^[9]. The debt was always there. The accounting was not.

Why the spend fails to convert is a question about how the tools behave. Large language models are trained to be helpful and generative, which means they are not optimized for exhaustive validation ^[13]. Asked to review a codebase, a model does not read every line; it searches for areas of interest and evaluates the context around them ^[14]. It leans toward returning the first good answer, and has to be explicitly told or configured to do more thorough work, which is where the iteration cost lands ^[15]. So a compliance-shaped question, of the form "does every element meet the standard," gets answered by a system that samples. More money buys more samples. It does not buy the guarantee.

That sets the sorting rule. Where the requirement is checking every element against a defined standard, every time, without exception, deterministic rules-based tools beat model calls on speed, cost, consistency and completeness ^[16]. The model earns its keep on the judgment calls, the synthesis, and the explanation of what a result means and how to fix it ^[17]. The criterion for which side a task belongs on is the requirement, not the department that owns it, and most AI programs have never run that sort.

The argument arrives from an interested party, which is worth saying plainly. It comes from the CTO of Deque Systems, writing from digital accessibility ^[1], a field he chooses as evidence because it has codified standards and measurable outcomes, so the difference between what AI does well and what it does badly shows up in numbers ^[19]. Take the recommendation with that in mind. The third-party figures he cites are checkable, and the description of model behaviour under a thoroughness requirement is not controversial.

His proposed remedy is unglamorous: an analytical framework built on consistent measurement of cost, speed and quality ^[20]. It reads like housekeeping until you remember what changed underneath it. AI is not priced by seat, it is priced by consumption ^[3], and the introductory rates were set to drive adoption rather than to reflect what serving the model actually cost ^[4]. Under seat pricing, an unmeasured task is a rounding error. Under consumption pricing, an unmeasured task is an open invoice, and the earliest adopters wrote their ROI cases in a currency that has since been repriced ^[2].