Start with the deck arithmetic, because it carries the whole thesis. A presentation that used to take a week now takes a day [12], and if you call a week five working days, the production cost per deck falls about 80 per cent; the same team then receives five times as many of them [12], nobody was hired to read them, and review hours therefore rise fivefold [17]. None of that appears on a vendor invoice. It appears as senior staff who cannot process what arrives, which is the decision paralysis one executive described to Webb [13], and it is what she means by CEOs buying abundance while nobody budgets for the cost of abundance [15].
The pilot side is the same shape with a bigger number attached. One of Webb's clients ran 14 or 15 generative AI and agent pilots since the start of the year, staffed by two-pizza teams, and scaled none of them [6], a conversion rate of zero on a denominator large enough to be a programme rather than a trial [18]. Her explanation is the load-bearing part: the pilots ran without legal and IT in the room, so each restarted from zero [7]. The fifteenth pilot inherited nothing from the first, so the money bought fifteen prototypes and no platform, and the integration work those pizza-fed teams were not doing is precisely the work that would have made any single pilot scalable.
Bain's June survey of 951 companies is the only measured figure in the pile: nearly 40 per cent of firms that tracked AI cost savings came in under 10 per cent against targets of 11 to 20 per cent [8], which is under half the top of the band they underwrote [19], and 90 per cent are raising budgets regardless [9]. Watch that denominator, though, since the 40 per cent is a share of the companies that bothered to measure, and the ones that did not measure are absent from it [20]. Webb's own evidence base is 100 to 150 CEO conversations a year [4], rich anecdote rather than a sample frame.
This is probably wrong in at least one direction, but the version I would defend is that the bust Webb sees in corporate AI spending [1] hits the buyer's own lines first, seats and consultants and pilot headcount, because the expensive part of her mechanism [2] is labour the buyer already owns and can cut inside a quarter, while compute contracts are dated and signed. The counter-thesis, which is the stronger one, is that those buyer budgets are the model builders' revenue, so a cut in corporate AI spending shows up in vendor top lines rather than beside them. A third reading sits under both: Marc Andreessen's March claim that large companies are overstaffed by as much as 75 per cent and use AI as a silver bullet excuse for pandemic-era corrections [10] fits awkwardly against Oxford Economics' finding that AI-cited layoffs were 4.5 per cent of US job losses [11], and if the cheap-production story were already converting into headcount, that 4.5 per cent would be larger [21].
What would falsify me is measurable. If a Bain follow-up shows the sub-10-per-cent cohort migrating into the 11 to 20 band while budgets keep climbing, the review burden was a transition expense and Webb's bust is a delay; if the number of firms she says have found sustainable experimentation, currently one hand's worth [14], starts needing a second, the mechanism is fixable rather than structural.
