The duplication described in the source column is ordinary rather than exotic. A familiar question comes back in slightly different wording, travel policy becoming reimbursement, then an exception request, then a manager asking what can be approved, and the system loads the same rules and pulls the same documents to arrive at roughly the same answer [7]. A single request like that barely registers. Multiply it across teams and agent workflows and it becomes a line on the operating bill [8].
The BCG figure is usually presented as AI spending doubling [1]. Put in dollars, the increase itself, 0.9 points of revenue, is larger than the entire current AI budget: this year adds more than everything spent to date [14]. That addition lands in a base where, taking the McKinsey survey at its word, fewer than one respondent in five reports a company-wide operating profit effect of 5% or more [15]. All of these figures reach us secondhand, through a Forbes Technology Council column by Erum Manzoor of BOX Motorsports and VectorE Ventures [13].
The McKinsey number could also reflect the age of these deployments rather than their waste: IBM's survey of CEOs put enterprise-wide scaling at 16%, against 25% of initiatives delivering the expected return, so much of the estate is still pre-scale [3]. The two readings do not compete, because they point at the same decision. If the thin profit signal is immaturity, the learning period is being funded at roughly twice last year's rate [14], and the cost of deferring architecture work until the pilots settle rises with it.
Klarna is the case worth keeping, because the savings were real and the accounting changed anyway. The company said in May 2024 that generative AI had taken about $10 million a year out of marketing [4]; by September 2025 its leadership said it had leaned too heavily on AI-led cost-cutting and moved attention back toward products, services and growth [5]. The trade-off is worth naming plainly: cost removed from one line can return as a revenue or service problem later, which means a saving counts only after the wider business has absorbed it.
The remedies are unglamorous. Prompt caching reuses identical instructions, semantic caching recognises the same intent asked in different words and returns an approved answer where that is safe, and the same layer decides which model handles a routine task and how many steps an agent is allowed to take [11]. OpenAI says caching can cut input token costs by up to 90% and time to first token by up to 80% on reused content [6]. That number comes from the party selling the tokens, and the words carrying the load are "up to", which is a ceiling rather than a planning figure. Reuse also creates an ownership problem, since somebody has to stay accountable for whether the approved answer is still approved.
The infrastructure figures in the same column describe a decade rather than a quarter. The IEA projects global data centre electricity use roughly doubling to about 945 terawatt-hours by 2030, with AI the largest driver [9], and McKinsey estimated in April 2025 that data centres could require about $6.7 trillion of capital spending over that period [10]. Those totals do not appear on any single company's invoice, but compute, memory and energy sit inside the business case whether or not they reach the slide [12]. The narrow decision available this quarter is who owns the reuse and routing layer, and whether that person also sees the bill. At 1.7% of revenue, someone in finance will ask.
