Bain & Company surveyed 951 large companies and found that AI cost savings are coming in well under what executives planned for. Nearly 40% of the firms that bothered to measure their savings landed below 10%, even though the most common target was 11% to 20%. The survey was completed in April and covers companies with more than $100 million in revenue across nine sectors.
Then there's the part that should worry shareholders. Despite the shortfall, 90% of those companies are planning to spend more next year.
The agents that aren't
The headline diagnosis getting passed around is that humans keep getting in the way of the algorithms. That's roughly true, but it flattens what Bain actually argues. According to the firm's Bain brief, only 7% of companies run fully autonomous agents in production. The most common setup, cited by 38%, still requires a human to approve every action. Another 32% run with guardrails, looping a person in when the agent hits something it can't handle.
The interesting bit is what Bain says about that. The report's authors don't call human oversight the mistake. They call it the right posture right now. The actual problem is narrower and more boring: the investment case got built on full-automation math, and the thing running in production routes a chunk of decisions to a human queue instead. The CFO signed off on one set of numbers. Operations is living with another.
"The technology worked. The value didn't arrive." That line from the report is doing a lot of work, and it lands, because it shifts the blame off the models and onto the org chart.
Companies that missed their targets were also the ones stuck at lower autonomy. Only 38% of the laggards had agents at guardrails level or above, versus 50% of the firms that hit their numbers. Make of that correlation what you will. Bain frames it as cause; it could just as easily be that the companies good at deployment are good at everything.
The data wall, again
The second barrier won't surprise anyone who has worked near an enterprise data team. 41% of respondents named data access and integration as the single biggest obstacle, ranking it above budget, skills gaps, and executive buy-in. A decade and hundreds of billions of dollars into data modernization, and companies still can't reliably reach their own data.
Here's the twist worth sitting with. The companies that delivered on their targets cited data as a bigger problem than the ones that fell short, 44% against 40%. Not because they're worse at it. Because they're deploying at scale and hitting the wall harder. The underperformers, meanwhile, complained more about budget and competing priorities, which Bain reads as a tell that AI never got real executive attention in those shops.
A circular bet
The funding math is where this gets uncomfortable. 44% of companies, the largest group, said they're paying for their next wave of generative and agentic AI out of savings from prior automation programs. The same savings that keep coming in below target. The report calls this "a circular bet with a structural leak," and that's the rare consultant phrase that actually clarifies rather than obscures.
One caveat on all of this: the savings figures come from companies measuring their own results and reporting them to a consulting firm that sells AI transformation work. The numbers are self-reported, the methodology behind each company's "savings" calculation isn't standardized, and a survey completed in April 2026 is capturing a market still mid-hype. Treat the precise percentages as directional.
Bain's prescription is organizational, not technical: audit what past automation actually returned before funding the next round, name someone accountable when an agent makes a bad call, and fix broken workflows before automating them. The report published in June 2026. Whether the 90% planning to spend more take any of that advice is the thing to watch when the next budget cycle lands.
