This is the topic on this site where the published numbers most obviously cannot all be true at once, and the collection is built around that rather than around a resolution of it. Vendors publish case studies with large returns. Analysts forecast that a large share of agentic AI projects will be cancelled. Adoption surveys report near-universal experimentation. All three can be simultaneously accurate only if most deployments are small, most benefits are modest, and most ambitious projects fail — which is roughly what the neutral data suggests.
The neutral data comes from Stanford's AI Index, which compiled organisational AI use rising to 78% of surveyed respondents in 2024 from 55% a year earlier, and regular generative AI use in at least one business function more than doubling from 33% to 71%. Among business functions, the ones nearest to industrial operations lead on reported cost savings: supply chain and inventory management at 61%, service operations at 58%. The compilation's own qualification is the sentence to keep: those cost and revenue effects are reported “most commonly at low levels”.
Widespread adoption producing small measured effects is not a contradiction of either the vendor case studies or the cancellation forecasts. It is what you would expect if the easy wins are real but bounded, and the ambitious autonomous deployments — the ones that would justify the word agentic — are mostly still failing to reach production.
What is missing is named. The most-quoted figure in this subject, an analyst prediction that over 40% of agentic AI projects will be cancelled, reaches these pages only at one remove: the original press release returned HTTP 403 to the fetcher and, on a later attempt, a human-verification challenge that this project will not attempt to defeat. Its poll, sample size and the related vendor ‘agent washing’ estimate are absent entirely, and no assertion here rests on the figure alone.