Gartner's $2.59 trillion AI forecast outruns the workflows meant to absorb it

The concrete version of the argument is a customer service transfer. A customer opens with an AI chatbot, gets escalated to a support agent, then to a specialist, and re-explains the problem at every handoff, because the model can hold context that the process has no way to pass along ^[10]. That is the whole claim in miniature. The constraint lives in the handoffs, and handoffs belong to whoever owns operations, not to whoever signed the licence.

The arithmetic is worth doing even though the two bases do not match. Cross-multiply McKinsey's value shares against Gartner's forecast and the algorithm layer maps to roughly $259bn of the $2.59 trillion, with people, culture and process mapping to about $1.81 trillion ^[16]. One figure is a share of value and the other a projection of spend, so this is a scale illustration and not an allocation. It does make the asymmetry legible: on McKinsey's split, 90 cents of value in every dollar sits in changes no purchase order can execute ^[5].

The incentive structure explains the sequencing. Jerry Haywood notes that the budget line is easy to approve because nobody lost a job backing the agreed next big thing ^[18], and that many leadership teams treat the deployment as procurement, learning the benefits, buying the licences and waiting for the result ^[8]. Redesigning a process carries the opposite risk profile: it is visible, and it has an owner who can be blamed for the disruption. So the spend clears quickly and the redesign queues, which is a fair description of how a doubled budget ^[2] produces incremental improvement.

The historical comparison in the piece can be read as a case for patience. When manufacturers swapped central steam engines for electric motors and left the surrounding infrastructure alone, the result was a productivity plateau lasting close to three decades ^[9], which would suggest early ROI misses say little about the technology itself. That holds, up to a point. The column attributes the plateau to new equipment inside old architecture and does not identify what ended it ^[17], so the analogy argues about this decade rather than this quarter. The narrower finding survives either way: Haywood reports client automation programmes that pass technical validation in full and then fail at the operational level ^[11].

The record behind the claim is thin. The 96% figure is vendor research ^[3], the MIT work is described only as widely discussed with no number attached ^[4], and the piece asserting that the technology already performs impressively and will keep improving ^[14] is written by someone selling into the gap it describes. Read the other way, the same survey leaves 4% of companies reporting dramatic gains ^[15], a number small enough that the question wording matters as much as the headline. That does not make the mechanism wrong, but it does mean a 2026 budget defended on this basis is resting on a single op-ed ^[6].

What the column does say about the firms that see real change is that they rebuilt their operating models around the tools ^[12]. A budget approved without that work scheduled and owned buys capability the organization has no route to spend ^[7].