The two headline percentages do not share an axis. Rockwell's figure counts manufacturers who have invested in AI or merely plan to [1], while McKinsey's counts those who have finished the work and embedded it in plant operations [2]. Subtract one from the other and you get 93 points [1], of which some unpublished share is the distance between an intention and a completed state rather than the distance between spending and results. Read as a straight execution failure, the pair overstates the problem by an amount the record does not quantify.
The figure that does connect them sits inside the McKinsey survey: close to 60% of manufacturers set no clear target for what an AI deployment should achieve [3]. That matches the failure Dimitar Dimitrov, founder of the software firm Accedia, describes when he asks clients which decision an investment changes and how they will know it worked, and finds that most cannot say [9][12]. A project without a target does not visibly fail, it simply never generates the evidence that would justify extending it past the pilot, which is roughly what the 2% is counting.
The arithmetic also limits how much the target-setting explanation can carry on its own. If close to 60% set no target, about 40% did [2]; hold the 2% against that group and, assuming every fully embedded operation came from it, at most one target-setter in twenty reached full embedding [3]. Writing the number down first is necessary and plainly nowhere near sufficient.
A skeptic will call the list of paying use cases circular. Visual inspection, predictive maintenance, production scheduling and energy optimisation [10] are exactly the four areas where a plant already keeps a metered number, so the finding reduces to a tautology: measurable things can be measured. That is fair, and the gate still earns its keep, because it is applied before the money moves rather than after. Quality control is where half of Rockwell's respondents began [4], which tells you the sequence follows the instrumentation.
The supporting numbers are thinner than their precision suggests. The 52% defect reduction comes from one site in the World Economic Forum's January 2026 Lighthouse update [5], not from a distribution, and the 16% average performance gain describes sites that paired AI with connected equipment and trained teams [6], which bundles capital spending and training into a figure that will get quoted as an AI return. Dimitrov sells AI and custom software development [12], and that does not make his test wrong, but buyers should notice the case for it rests on reasoning rather than on returns anyone has published.
The trade-off inside the discipline is worth naming plainly. Funding only what finance already tracks buys a defensible business case and hands the roadmap to whatever the dashboard happens to count: energy per unit clears review quickly because it is already a line item, where a three or five point drop can show up the same month [14], while an inspection standard drifting between plants in different countries is a real cost with no meter on it [13]. So the choice an operations budget makes this quarter is between the small provable win and the instrumentation that would let a larger claim be proved next year, and with 78% of executives planning to route a fifth or more of improvement budgets into smart manufacturing [7], scope discipline rather than available money decides which one they end up with.
