Bain says 40% of companies are below their AI savings targets. IBM says 75% of AI projects fail to meet ROI expectations. NBER says 90% of companies report no measurable productivity impact from AI. These aren't small surveys — and they all landed in the same week.
The Problem
The "spray AI everywhere" era is officially over. Three major studies published this week paint a remarkably consistent picture: enterprises are spending more on AI than ever while getting less return than they expected.
Here's the breakdown:
- Bain & Company: 40% of companies are falling short of their AI cost savings targets. Not missing by a little — by enough that budgets are being questioned at the board level.
- IBM Institute for Business Value: 75% of AI projects fail to meet their stated ROI expectations. Three out of four projects. That's not a batting average any CFO would accept.
- National Bureau of Economic Research (NBER): 90% of companies report no measurable productivity impact from their AI investments. Nine out of ten.
And then there's the Uber story. They exhausted their annual AI coding tool budget by April. Four months into the year. Microsoft is canceling Claude Code licenses due to cost. JPMorgan allocated a $19.8 billion tech budget with 2,000 AI staff — and agents are now considered "core infrastructure," which means the spend is permanent.
The money is flowing. The returns are not.
The Problem Is Structural, Not Technological
This isn't about AI being overhyped. The technology works. The problem is how organizations deploy it.
The demo-to-production gap. AI tools look magical in controlled demos. In production, they hit edge cases, require constant maintenance, and produce inconsistent results. The 37% gap between lab benchmarks and real-world performance isn't a model problem — it's a deployment problem.
Vibe coding hangover. Developers felt 24% faster using AI coding tools, according to the METR study. Actual productivity? 19% slower. The perception-reality gap is enormous. Teams are generating more code but spending more time fixing it, reviewing it, and cleaning up inconsistent patterns.
Cost opacity. Enterprise AI spending is opaque by design. Token-based billing hides what flat-fee subscriptions used to make visible. When your annual budget burns in four months and you can't clearly attribute costs to outcomes, you have a financial control problem, not a technology problem.
Misaligned incentives. Vendors sell AI as a cost reducer. Internal teams buy AI as a capability multiplier. Neither party is measuring the same thing, so ROI calculations are fiction before the project starts.
What Actually Works
The 10-25% of projects that do hit their ROI targets share common patterns:
Start with a specific, measurable problem. Not "let's use AI to transform customer service." Instead: "reduce average ticket resolution time from 4 hours to 1 hour." The narrower the target, the easier it is to measure ROI.
Build short pipelines. As the reliability math shows, 5-step AI workflows at 95% reliability (77% success) dramatically outperform 20-step workflows (36% success). Simpler systems are more reliable, cheaper to run, and easier to measure.
Instrument everything from day one. Per-step token costs, success rates, time-to-resolution, and human escalation rates. If you can't measure ROI at the individual workflow level, you're flying blind at the portfolio level.
Budget for maintenance, not just deployment. Model upgrades break things. Agent behavior drifts. The cost of keeping an AI system running is 3-5x the cost of building it in the first year.
Benchmarks
The numbers that should inform every AI investment decision:
- 75% of AI projects fail ROI (IBM) — the base rate. Plan around this.
- 90% report no measurable productivity gain (NBER) — if you can't measure it, it doesn't exist.
- 40% below savings targets (Bain) — even when targets exist, they're usually missed.
- AI hardware costs up 4x in 6 months — the cost side of the ROI equation is getting worse, not better.
- 19% slower developer productivity despite feeling 24% faster (METR) — perception lies.
- JPMorgan: $19.8B tech budget with AI as core infrastructure — the scale of serious enterprise commitment.
- Uber: annual AI budget exhausted by April — the scale of serious enterprise overshoot.
Caveat: These studies measure large enterprise deployments. Smaller companies with more focused AI use cases may see better returns. The failure rate is highest in "transformation" projects and lowest in targeted automation of specific workflows.
Impact
For every $1 million spent on AI projects this year, roughly $750,000 will generate no measurable return based on current failure rates. Scale that to JPMorgan-level budgets and you're talking about billions in wasted investment across the enterprise landscape.
But here's the real cost: organizational fatigue. Every failed AI project makes the next one harder to justify. AI teams that burned through budget without delivering results don't get second chances — they get layoffs. The NBER finding that 90% of companies see no productivity impact isn't just a statistic. It's a future where AI investment freezes because early bets didn't pan out.
The compute cost crisis makes this worse. If hardware costs 4x more than six months ago, the ROI bar just got four times higher. Projects that were marginal at old prices are now clearly unprofitable.
The Companies That Win Won't Be the Ones That Spend the Most
The AI investment story of 2026 won't be about who spent the most. It'll be about who measured the best.
The 25% of projects hitting ROI targets aren't luckier — they're more disciplined. They start with measurable outcomes. They build simple systems. They track costs obsessively. And they kill projects that aren't performing instead of doubling down.
If your AI strategy doesn't have a kill criteria, it's not a strategy. It's a prayer.