Where AI ROI Hides
Two findings from the same year, 2025:
One: MIT's Project NANDA studied 300 enterprise AI deployments and concluded that about 95% of generative-AI pilots produced no measurable P&L impact. The stat rattled markets and launched a thousand LinkedIn posts.
Two: in the same organisations, employees privately reported the opposite — the same MIT work found a thriving “shadow AI economy,” with workers at the majority of companies using personal chatbot accounts for real work, and swearing by the hours saved.
Everyone's more productive. The company sees nothing. Both findings survive scrutiny. Reconcile them.
Sit in the contradiction. Real hours saved at desks; zero showing up in the P&L. Where, physically, between a person's saved hour and a company's income statement, could the value be leaking out?
What's the main reason the pilots showed nothing, per the research?
Pick one — committing first is what makes the answer stick.
the lesson continues after you choose
The headline reading — “AI doesn't deliver ROI” — is the one the market briefly panicked over. It's reasonable: 95% is a damning number.
But it can't survive the second finding: value was visibly being created at the desks of the same companies. A number that damning, next to behaviour that enthusiastic, isn't a verdict on the technology. It's a map of where value leaks between a saved hour and a financial statement — and the leak has three specific holes.
Hole one: the wrong tasks. Failed pilots clustered in flashy, client-facing, loosely-specified work — exactly the quadrants the CoReCo map says to avoid. The MIT data adds a twist most boards get backwards: over half of AI budgets chased sales and marketing, while the clearest measured ROI came from unglamorous back-office automation — the clear-steps, high-recurrence bubbles. ROI lives where work repeats and success is checkable, not where demos impress.
Hole two: verification cost. The honest equation is net value = time saved − time verifying − rework when verification was skipped. Fluent output makes the first term look huge and hides the second — until unchecked AI work lands on a colleague's desk. Researchers coined a word for that landing: “workslop” — polished-looking output that transfers the real effort downstream to whoever must decode or redo it. A task with cheap verification (code that runs, a reconciliation that balances) keeps its ROI; a task where checking costs as much as doing never had any.
Hole three: gains that never reach a bottleneck. An hour saved only becomes P&L when it turns into more throughput, fewer errors, or lower cost at a constraint. Scattered individual time-savings — the shadow economy's kind — evaporate into slightly longer coffee breaks unless the workflow around them is redesigned to bank the gain. That's why individuals feel rich while the company measures nothing: the value is real, unbanked, and invisible to systems that never set a baseline.
So the contradiction was never a contradiction. The 95% and the shadow economy are the same fact seen from two floors of the building: individuals capture AI value task by task, informally, instantly; organisations only capture it through task selection, integration and measurement — the boring machinery most pilots skipped. The stat doesn't say AI lacks ROI. It says ROI hides in the back office, inside recurring tasks, behind a baseline someone bothered to measure — exactly where the CoReCo map has been pointing all along.
Your rule — the demo test: before greenlighting any AI initiative, require three sentences: the metric it moves, the owner who answers for it, and the workflow step it replaces. Can't produce all three? It's a demo, not a pilot — run it as one, cheaply, and don't book its ROI. Then pick your real pilot off the CoReCo map, where the three sentences write themselves.