Three sources point at the same underlying mechanism from different disciplines, and read together, they explain why the gap persists rather than closing as AI matures.
Accounting theory has a name for part of this problem. Corrado, Hulten, and Sichel's foundational work on intangible capital found that national accounts have historically excluded hundreds of billions of dollars of intangible investment, in software, organizational process, and knowledge capital, because it does not fit categories built for physical plant and equipment (Corrado et al., 2009). Enterprise accounting inherits the same categories. Under IAS 38, the international standard governing intangible assets, costs incurred in the research phase of an internally generated intangible must be expensed as incurred, and only a narrow band of development costs meeting strict criteria can be recognized as an asset (IFRS Foundation, n.d.). Most enterprise AI investment, prompt engineering, data preparation, workflow redesign, model fine-tuning, falls on the expensed side of that line. It reduces reported profit in the period it is incurred and creates no balance-sheet asset against which a later return can be measured. The absence is not a reporting failure. It is the standard working as designed, for a different kind of economy.
Brynjolfsson, Rock, and Syverson's Productivity J-Curve supplies the mechanism's time dimension. Their model of general-purpose technologies shows that early adoption periods systematically understate output and productivity, because firms are investing heavily in unmeasured complementary intangibles, retraining, process redesign, new organizational routines, that do not show up as capital in the accounts. Only later, once those investments mature into visible output, does measured productivity catch up, and can even overshoot (Brynjolfsson et al., 2021). Applied to the present pattern, this suggests at least part of today's invisible AI return is not permanently invisible. It is temporarily unmeasured, and the accounting system is not built to distinguish a temporary lag from a permanent absence.
That distinction is the analytical contribution this note adds to the evidence. The research reviewed here typically treats "AI investment isn't showing a return" as one condition. It is more useful, and more actionable, to treat it as two different conditions that look identical on a P&L: measurement lag, where real value exists but has not yet compounded into recognized output, and measurement blindness, where the value form itself, non-rival reuse of a model across business units, compounding accuracy gains, avoided downside risk, will never register cleanly in accounts built around discrete, depreciable, single-use assets. Confusing the two produces exactly the pattern Gartner documents: some pilots reporting no return are cancelled for underperforming when they are, in fact, still in the lag phase; others are kept running on faith when their value is structurally blind and no future accounting cycle will resolve the question.
Read together, the three research streams describing organizations that do show results, MIT NANDA's roughly 5 percent extracting real value, Deloitte's one-in-five ROI leaders, and McKinsey's workflow-redesign cohort, are plausibly describing an overlapping population: firms that have built parallel, purpose-built measurement (reuse rate, capability trajectory, cycle time) alongside standard accounting, rather than waiting for GAAP or IFRS categories to catch up. The redesign is not incidental to their results. On this reading, it is close to a precondition for them.
