Why the same ROI number means opposite things

Two numbers circulate more than any others in AI-ROI coverage this year: a claimed $3.70 return per dollar of AI spend, and a report that 95% of enterprise generative-AI pilots fail to reach P&L impact. Cited next to each other, they read as a contradiction. They are not — they are measuring different things.

The $3.70 figure comes from an IDC survey of 2,000 enterprises, commissioned by Microsoft. It is a self-reported, modeled multiple across companies that have already committed budget to AI programs — survivorship built into the sample before the first response is recorded. The 95% figure comes from MIT research on a broader set of pilots, most of which never scaled past initial deployment. One measures return among the AI programs that made it far enough to have a return to report. The other measures failure among all the programs that were tried.

Neither number is wrong. Both are frequently cited without the cohort attached, which is the actual problem: a board that hears "$3.70 return" assumes it applies to their next pilot, when it describes only the subset of pilots that already worked. This site's measure taxonomy — Adoption, Cost, Outcome — exists specifically to stop that kind of unattributed transfer. Before citing either number, ask what population it was measured against, and whether your program looks like that population yet.

The IDC figure in particular deserves more scrutiny than it usually gets: it is vendor-sponsored, and nobody in general circulation appears to have fact-checked it against its actual survey methodology. That is not a reason to discard it — it is a reason to label it, the way this site's vendor and research entries carry a verification date, rather than letting it travel as an unattributed constant.