There's a survey making the rounds this week, and the number that stuck with me isn't the disappointing one. It's the dangerous one.
Bain & Company ran their Automation and AI Pathfinder Survey in April 2026. They asked 951 companies, all of them with more than $100 million in revenue, how their AI spending actually paid off. The results came out June 1 through Bloomberg.
Only 4% of those companies hit more than 30% in cost savings from AI. About 40% got 10% or less. A big chunk, 37%, had told themselves they'd see cuts of 10 to 20%. Most of them landed in the single digits.
Bain put it cleanly: the technology worked. The value didn't arrive.
That's the disappointing number. Here's the dangerous one.
The bet most people don't realize they're making
44% of those same companies are funding their next round of AI investment using the savings from the last round. The savings that, for most of them, never showed up.
Read that again. They're paying for wave two with money wave one was supposed to generate but didn't. And roughly 90% are increasing their AI budgets anyway.
Bain called it "a circular bet with a structural leak." Their reasoning is exactly right. The prior wave underdelivered, so the savings pool is smaller than assumed. And the business case for the new wave was sized against projections instead of actuals. You're not reinvesting profit. You're reinvesting a forecast.
When you're a billion-dollar company, you can eat that. A phantom-ROI miss is a rounding error and a slightly awkward board slide. You absorb it and move on.
When you run a small business, a missed forecast you already spent isn't a rounding error. It's payroll. It's the quarter.
Why the savings didn't show up
Here's the part nobody wants to hear, because it's not as fun as blaming the robots.
The number one reason AI programs underperformed wasn't budget. It wasn't strategy. It wasn't the model. 41% of companies said the biggest barrier was data access and integration. They couldn't reliably get at their own data to feed the thing they bought.
There's a second piece. Most of those investment cases assumed full automation economics, where the AI just runs the whole job start to finish. But only 7% of these companies actually run fully autonomous agents in production. They budgeted for a level of automation that mostly doesn't exist yet, then acted surprised when the math came up short.
So the gap isn't really about AI failing. It's about measuring the wrong thing, at the wrong time, against a number nobody banked. Researchers at UC Berkeley made this exact point about an earlier round of grim AI studies: a lot of what gets reported as "AI failure" is actually measurement failure. Which, honestly, should make you feel better, not worse. The fix for a measurement problem is in your control.
The small business version of a trillion-dollar mistake
This same pattern is happening at every size right now. The whole AI-bubble conversation is about hyperscalers spending hundreds of billions against a fraction of that in real revenue, financing the future on returns that haven't landed. The Bain survey is the corporate-budget version. And there's a small business version too, even if it never makes Bloomberg.
About 68% of US small businesses now use AI regularly, up from roughly half in mid-2024. That's real adoption. But by most estimates, 77% have no written AI policy, and only somewhere around 15 to 20% use it strategically, meaning they picked a specific workflow, trained the team, and actually measure outcomes instead of activity.
The savings math is also worse for you than it looks on the pricing page. Industry analysis pegs the sticker price of an AI tool at maybe 50 to 65% of what it really costs once you add training time, usage overages, and the work of bending it into how you actually operate. Treat those as estimates, not gospel. But the direction is obvious to anyone who's done it. The number on the website is the smallest number you'll pay.
And the savings can go negative. There are companies right now that cut staff for AI and are quietly rehiring, because the work didn't get done. Multiple surveys this spring found a meaningful share of those firms spent more rehiring than they ever saved. That's not an AI problem. That's a "we budgeted on a number we never measured" problem.
What actually works
The companies in Bain's survey that came out ahead did one boring thing. They treated data and process as a leadership problem, fixed it, and funded the next wave from returns that actually materialized. Real money, banked first.
You can do the small-business version of that, and it's genuinely simpler.
Pick one workflow. Not a department, not a "transformation." One workflow, in whatever part of the business is closest to the money, usually customer service or marketing or whatever's eating your week. Wire automation to that single thing.
Then run it for about 90 days and measure three things, honestly:
- Time saved per week, counted, not guessed.
- Output quality compared to how you did it before. Same or better, or it doesn't count.
- Total cost. Subscription, plus training, plus the disruption while everyone got used to it.
If after 90 days you've actually banked savings, then you've got real money to decide what to automate next. If you haven't, you just saved yourself from funding round two on a fantasy. Either outcome is a win, because both are based on something that happened instead of something you hoped would.
The trap isn't AI. AI's fine. The trap is letting a forecast act like income.
We do this wiring and measurement for clients. Picking the one workflow that moves money, instrumenting it so you can see what it actually saved, and only then deciding what's worth building next. That's the part that turns AI spend into AI return, and it's the part most people skip.
If you're sizing up your own AI budget and you're not sure which number is real and which one is a hope, that's worth a conversation. The first one's free, no commitment. Or grab a free membership at kief.studio for the companion resource that walks through the 90-day measurement setup.