The number has been out there for months. Ninety-five percent of enterprise AI projects fail to deliver measurable return on investment. Every consultant in the room has a theory: the AI was hallucinating, the data was dirty, the change management was weak.
Nobody is asking the obvious question: what if we're measuring the wrong thing?
The measurement trap
Arvind Narayanan, who has spent the last two years systematically debunking AI hype at Princeton, makes a point that most executive presentations ignore: AI is a normal technology. It doesn't create overnight revolutions. It diffuses slowly, unevenly, and its impact follows the same S-curve every general-purpose technology has followed since the telegraph. The companies measuring AI impact after six months of deployment are asking a question with no good answer yet.
But there's something worse than measuring too early. It's measuring the wrong thing entirely.
Most enterprise AI projects are built to make existing processes faster. Customer service automation that handles tickets the same way, just quicker. Document review that follows the same checklist, just without the associate's billing hours. Report generation that produces the same slides, just overnight instead of by Thursday.
The ROI question companies ask: did this task get cheaper? The question they should ask: should this task exist at all?
AI doesn't make a broken process more efficient. It makes it fail faster, and at scale.
The 19% problem
In July 2025, METR published the results of a randomized controlled trial on developer productivity. The finding was uncomfortable: experienced developers using Cursor and Claude tools took 19% longer to complete tasks than when working without AI. More importantly, those same developers estimated they were 20% faster before seeing the results.
This isn't a story about AI being useless. It's a story about the gap between perceived and actual value, compounded at every level of an organization. Individual contributors think the AI helps. Managers believe the productivity reports. CFOs approve the renewals. And somewhere in the middle, the actual work takes longer than before.
The one developer who genuinely got faster, by 38%, had logged more than 50 hours of deliberate practice with the tool. Not watching a vendor demo. Not a pilot program. Fifty hours of real use, failure, and iteration, on projects that actually mattered.
That detail reveals the structure of the problem. AI tools, like any sophisticated instrument, require investment that doesn't show up in a quarterly review.
Wrong question, wrong answer
Forrester's 2026 data shows that 25% of planned AI spend is being deferred. Only 15% of AI decision-makers have seen a positive impact on profitability. The narrative is shifting, quietly, from "how do we adopt AI" to "how do we justify what we've already spent."
This is actually a useful moment.
The 95% failure rate isn't evidence that AI doesn't work. It's evidence that companies are optimizing workflows that should be redesigned, not accelerated. They're applying a fast engine to the wrong chassis and measuring the speed of the exhaust.
The organizations that will find real returns aren't the ones with the biggest model budgets. They're the ones willing to ask which processes shouldn't exist, and use AI to eliminate them entirely, rather than run them faster.
Why this matters for your business
That's not a technology question. It's an organizational design question, and AI has made it urgent in a way that incremental improvement never could.
When I work with companies on AI integration, the first conversation is never about which model to use. It's about which meetings, which reports, which approval chains, which review cycles exist only because removing them would require someone to take responsibility for the decision. AI gives organizations an excuse to ask that question without it feeling like an attack.
Use the excuse. The ROI question will answer itself.