Seventy-three percent of executives are disappointed with their AI investments. Seventy percent are ready to slash budgets. And 69% say their people spend more time monitoring AI than doing the original work the AI was supposed to handle.
This isn't a startup complaining on Twitter. This is a global enterprise survey from G-P, and the numbers are brutal.
The Problem: AI ROI is a Mirage for Most Companies
The enterprise AI adoption story was supposed to be simple: buy tools, save time, cut costs, grow revenue. That's the pitch every vendor makes. That's the slide in every board deck.
Reality looks different. The G-P report found that only 29% of executives see significant ROI from generative AI. The rest are somewhere between "meh" and "we burned cash." Worse — 95% of AI pilots deliver zero measurable P&L impact.
The killer stat: 69% of teams say humans now spend more time supervising AI than they did doing the work themselves. This is the AI version of buying a robot vacuum that requires you to pre-clean the floor before it runs.
And it's not just small companies feeling the pain. Microsoft is canceling Claude Code licenses. Uber burned through its entire 2026 AI budget in four months. Nvidia's own VP admitted compute costs now exceed employee costs at some organizations.
The Solution: What the 27% Actually Did
The companies seeing real returns aren't using better models. They're using better systems. Here's what separates them:
They started with infrastructure, not models. The top frustration — 88% of agent failures — comes from infrastructure gaps, not model quality. Context blindness, rogue actions, and silent degradation kill more projects than bad prompts. The winners fixed their plumbing first.
They measured obsessively. Not vanity metrics like "number of AI interactions." The 27% tracked time-to-completion, error rates, and actual cost per task — including the hidden cost of human supervision.
They avoided vibe-driven adoption. There's a reason "vibe coding" is becoming a pejorative. Developers who spent 30 weekends building with AI are now spending weekends rewriting what they built. The winners used AI for bounded, testable tasks — not open-ended creation.
The Benchmarks: What Real ROI Looks Like
Here's the uncomfortable honesty — benchmarks for enterprise AI ROI are mostly self-reported and inconsistently measured. But the data we do have points to a pattern:
- Cognition (Devin) hit $492M ARR with 50% month-over-month growth — but they're a pure AI product company, not a traditional enterprise adopting AI internally
- Cost savings in the 27% camp typically range from 15-30% on targeted workflows (customer support, document processing), not the 50-80% vendors promise
- Time savings are real but frequently offset by supervision overhead — net savings often land at 10-20% after accounting for monitoring time
- Implementation timelines for the successful 27% averaged 6-9 months from pilot to measurable P&L impact — not the "deploy in a weekend" narrative
Caveat: these numbers come from vendor reports and selective case studies. The full picture of the 27% is still emerging because most companies doing it well aren't publishing their playbooks.
The Impact: Why This Matters Right Now
The AI investment backlash isn't theoretical — it's actively reshaping budgets. Seventy percent of executives are ready to cut AI spending. That means the window for demonstrating value is closing.
For businesses evaluating AI right now, the calculus has changed:
- Don't adopt AI to adopt AI. Every tool must answer: what specific task does this replace or augment, and how do we measure it?
- Budget for supervision. If you're not accounting for human oversight costs, your ROI model is wrong. Period.
- Start narrow. The companies winning with AI didn't transform their entire operation. They picked one painful workflow and automated it well.
- Watch the token pricing crisis. If Microsoft and Uber can't make the economics work, smaller companies need to be especially careful with metered AI tools.
The total enterprise AI market is still growing — Anthropic just raised $30B at a $900B+ valuation. But the gap between investment and returns is becoming impossible to ignore.
The Bottom Line
The AI industry has a confidence problem, and it's self-inflicted. Vendors overpromised. Enterprises under-prepared. And now 73% of the people writing the checks are wondering if they got sold a bill of goods.
The technology works. The deployment strategies mostly don't. If you're building or buying AI tools in 2026, focus less on which model to use and more on whether your organization can actually absorb, monitor, and measure the impact. The 27% figured that out. The other 73% is still waiting for the magic to happen on its own.
It won't.