The four figures that get mixed up
| Figure | What was actually measured | Source & year |
|---|---|---|
| 95% fail | No measurable impact on the profit and loss statement from generative AI pilots. Not “stopped,” not “technically failed.” | MIT NANDA, 2025 |
| 30% discontinued | A forecast by the end of 2025, no measurement. Later revised to at least 50%. | Gartner, 2024/2025 |
| 26% achieve value | Self-assessment of a thousand managers across thirty AI skills; 49% are stuck in pilot projects. | BCG, 2024 |
| 74% achieve the ROI | Only the most advanced initiative that has already been scaled up at each organization. | Deloitte, 2025 |
About that 95 percent. The figure has been heavily criticized: the evidence has not been published, and the issuing party sells agent infrastructure that benefits from the conclusion that current tools fall short. Never use it as standalone evidence. What holds up is the direction all the sources point to: broad adoption, rare transformation.
The 10-20-70 rule
BCG asked a thousand managers in 59 countries where their problems come from. The breakdown: approximately 10 percent is in the algorithms, 20 percent in technology and data, and 70 percent in people and processes. Companies that do extract value distribute their resources exactly this way.
CBS asked the question from the other angle. Among companies that considered AI but didn't get started, the main reason for more than seven in ten was lack of experience. Two different studies, two different populations, the same answer.
The practical translation. If 70 percent of your risk lies in people and processes, the most important question isn't "which tool" but "which task, who does it now, and where is how it works documented?" That's exactly the order we follow.
The way it works now
- The appointment is in one colleague's mailbox
- Data is manually retyped from one system to another
- The person who knows is on vacation
- No one can see afterwards why this decision was made
As it is
- The agreement is stored with the client, findable for everyone
- Once entered, information flows through to the rest
- Work continues, even if someone is absent
- Every decision has a source, an owner, and a date
What the winners do differently
This has happened before
In 1987, economist Robert Solow noted that you could see the computer age everywhere except in the productivity figures. Paul David explained in 1990 why: around 1900, the electric motor still powered less than 5 percent of factory machinery, and only in the twenties — when factories completely reorganized their layout — did the productivity gains come.
So the technology was there decades before the returns. Not because the technology was no good, but because the work itself hadn’t yet been redesigned. That’s exactly what BCG’s 70 percent says, a hundred years later.
First find out what it costs you
Before you automate something, it's useful to know where it's leaking. Eight questions, ten minutes, and you get your own number back.
All figures on this page have been checked against the original publication. Where a figure is a forecast rather than a measurement, or where the methodology has been criticized, this is noted.
Read more: AI use in SMEs: the figures and preserve knowledge when employees retire.