Executive summary
78%
of companies use AI regularly, per McKinsey (Nov 2025)
33%
are scaling it across the organisation, in the same survey
~5%
of pilots reach lasting financial gain, per MIT NANDA (a contested, non-peer-reviewed figure)
Key conclusions
- Almost every company now uses AI, but few scale it. Both groups use the same technology, so the difference lies in the systems and practices around it.
- Four things separate scaled deployments from stalled pilots: a use case chosen for real operational friction, data available on production terms, a baseline set before go-live and a named owner for month thirteen.
- A stalled pilot leaves behind an unmaintained connector and a copy of the data that has drifted from the source. These leftovers make a second attempt more expensive than the first.
Getting an organisation to try AI is no longer the hard part. The value is captured or lost in the step from trying AI to running it across the business.
The two McKinsey figures describe the gap from both ends: almost every company has adopted AI, but few have scaled it. The MIT NANDA figure is often quoted alongside them, and it needs a caveat. It comes from a small sample (on the order of 150 survey responses and 50 interviews) and was not peer reviewed. It also counted a pilot as failed if it showed no measurable P&L impact within six months. That definition has drawn substantive methodological criticism since publication. We cite the figure only as a directional indicator that agrees with the larger survey. The argument below rests on the McKinsey gap and holds without NANDA's number.
Both groups use the same technology, so the difference must lie in what surrounds it.
Four practices separate scaled deployments from stalled pilots
| Stuck at pilot | Scaled |
|---|---|
| Use case chosen for how impressive it demonstrates | Use case chosen for where the operational friction actually is |
| Data assembled specially for the pilot | Data available on the same terms the production system will have |
| Success measured against the demo | Success measured against a baseline established before go-live |
| Ownership ends with the project team | Ownership named for month thirteen, and priced |
Each stalled pilot makes the next attempt more expensive
Every stalled pilot leaves behind a partly built integration: a connector that nobody maintains, a copy of the data that has started to drift from the source, and a process with a manual step added back to work around it.
The next project has to work through these leftovers, which is why a second attempt usually costs more than the first.
