Median quarterly AI spend of roughly $1,500 rising to nearly $44,000 is about a 29-fold increase [9], and in technology firms specifically DX puts the rise at roughly 28 times [4]. Over the same four quarters, DX found the innovation ratio, meaning the share of engineering effort going to new capability rather than maintenance, essentially flat [5]. Whatever the money bought, it did not visibly move the mix of work the engineering organisation performs.
The mechanism Taplin describes is not deception but incentive alignment that stops at the launch date. An initiative is approved on a demonstration, built, shipped, announced, and then attention moves on with nobody assigned to establish whether it worked [6]. He argues every incentive is satisfied at the announcement and none extends past it, so the organisation learns that the announcement is the deliverable [7]. That is a governance failure with a specific location: the approval gate has an owner, and the verification gate does not.
The measurement problem compounds it. Taplin's contention is that a typical AI dashboard tracks model performance, usage and milestone completion, none of which is a business metric, and all three can improve while value declines [11]. Rising usage in particular is ambiguous, since people use a system more when they trust it and also when they have to check its work [10]. DX's own instruments diverged for a year: its Developer Experience Index fell from 67 to 65 across four quarters while measured output rose [12]. When two instruments pointed at the same organisations disagree, the one on the executive dashboard is the one that reported improvement [13].
One objection to the 89-to-six gap [2][3] is that the six percent figure is a survey artefact. Executives are asked whether they are "sure" they have clear examples of organisation-wide ROI, which is a high bar, and 173 executives is a small sample [1]. That objection may explain the precise level, but it does not change the direction. Even generously read, the gap says confidence in AI's effect is widely held and evidence for it is not, and the spend figures say the budget decisions did not wait for the evidence [9].
Why capable organisations with boards, auditors and skeptical CFOs miss this is answered, at least in part, by the RAND interviews of 65 practitioners with five or more years building AI and ML models, which found the most common root cause of failure was business leaders not knowing how to set a project up to succeed [8]. Taplin's own read from more than 200 recorded conversations with CTOs and founders is that failures were rarely about model quality, but about nobody defining what "working" meant and nobody accountable for finding out [14]. Both point toward a definitional fix rather than a technical one.
The tradeoff for anyone deciding this quarter is between speed of approval and defensibility of approval. Requiring a named baseline metric and a named owner for post-launch verification will slow the intake of AI proposals, and some genuinely good ones will die in the queue. The cost of not requiring it is that next year's budget conversation happens with a spend base roughly thirty times larger [9] and the same six percent evidence base [3]. Which of those two costs is cheaper depends on how much of the current portfolio is real, and this record does not tell us that yet.
