The four uses in the works column share a property that matters more than the labels: the buyer can check the output within the week. Code either passes review or it does not [4], and an AI voice system either completes the rudimentary action or the caller rings back [6]. Security sits slightly further out, since a detection platform's value shows up as alerts triaged rather than incidents avoided [5], but even there the feedback arrives in days. That is the real inheritance from enterprise spending: a set of use cases where a firm with no research budget can tell, quickly and from its own numbers, whether it is being sold something that works, rather than a vendor list.
What the free R&D does not include is unit economics. The column reports that large firms discovered astronomical compute costs once employees and agents began consuming tokens at volume [8], and Marks says smaller firms have watched large projects burn billions with little to show for it [9]. Neither observation tells a ten-person company what its own monthly bill will be at its own usage. Marks's answer is abstention and low-hanging fruit [9], which is a reasonable posture and not a forecast. On SMB token spend, the record here says nothing, and pretending otherwise would be the kind of borrowing that gets expensive.
The board-deck version is tidy: four cleared uses and three known failure patterns, plus one columnist [1]. An operator's account is not evidence on measurement, and that criticism holds, because nothing in this column is counted. The claim worth keeping is narrower. It is that the cost of verifying an AI purchase has fallen unevenly across categories, and the categories where it has fallen furthest are the ones large budgets have already stress-tested.
Then there is the sequencing trade, which is the part a shortlist hides. Deferring agents is close to free while their reliability is the thing in doubt [13], and Marks is content to let large companies beta-test them [14], a patience he supports by noting that even Sam Altman has called OpenAI's adoption timelines too ambitious [16]. The cost lands later. The firms running unreliable agents now are also the firms accumulating review habits and failure taxonomies, and the deferral offers no trigger for revisiting the decision. This quarter that trade is obviously right. The quarter agents start holding up, whenever that is, it converts into a standing start.
The one lesson with no capital cost is the communications one. Marks argues that public enthusiasm for headcount reduction on earnings calls produced backlash [10], and that smaller employers have taken the opposite line, using AI for productivity while leaving payroll alone and recruiting the people that policy unsettles [11]. He goes further, saying corporations used AI as cover for over-hiring and then rehired the same staff [12]. That is the hardest claim in the piece to verify and the cheapest to act on, which is an awkward combination. A small firm that adopts it is making a bet on labour supply, not on a technology, and it should price it that way.
