Bain surveyed 951 companies with over $100 million in revenue across nine sectors and found that among those tracking their AI cost savings, the largest group, 40 percent, saw reductions of 10 percent or less. That is far below what most executives expected when they approved the spending. Worse, 44 percent of large companies funding their next round of AI investment are basing that spending on savings from the last round, savings that have not actually shown up yet. Bain called this a circular bet with a structural leak rather than genuine discipline.
The report identifies the core bottleneck as access, not ambition. Despite hundreds of billions spent globally on data modernization, the top reason AI programs underperform is that companies still cannot reliably get to their own data. Bain's fix is to stop waiting for perfectly structured data and instead start feeding AI whatever is usable now, then use AI itself to help sort the rest.
What's underneath it: this is a sequencing problem more than a technology problem. Organizations bought the tool before they solved the access problem, then measured the tool against a projection instead of against what it actually returned. That gap between projected and actual value is exactly what a companion Gartner report points to as well, predicting that over 40 percent of agentic AI projects will be canceled by the end of 2027 due to unclear business value and cost overruns. Gartner's read is that most agentic projects right now are hype-driven proofs of concept rather than production systems built around a specific, measurable job.
The takeaway for anyone running their own automation, not just enterprise IT departments, is that the value of an AI pipeline has to be measured against what it actually saved or produced, not against what the pitch deck promised. A tool that works technically but never gets validated against real before-and-after numbers is running on hope, not on data.