The technology is rarely what stalls. Manufacturers run pilots that work, and then cannot get past the second site, because a pilot needs a champion and a rollout needs an owner, a measure and a rhythm that crosses plants.
The applications in this category are well established. The failure is almost never in the model.
The four places it genuinely works
Predictive maintenance. The most proven of the four. It requires sensor history you may not have yet, and when you do not, the real project is instrumentation, which is slower and duller and has to come first.
Visual quality inspection. Mature, and the economics are usually clear because the defect cost is already known.
Demand and inventory forecasting. Valuable and dependent on the quality of your historical data, which in most businesses is worse than anybody wants to admit until the project starts.
Scheduling. Real gains, and the hardest to adopt, because the schedule carries a great deal of undocumented knowledge about which jobs genuinely cannot follow each other.
Where it actually stalls
| Stage | What happens |
|---|---|
| Pilot | Works. A champion drives it at one site |
| Second site | Different equipment, different shift patterns, different people |
| Scale | Needs IT, operations, finance and plant leadership to move together |
| Steady state | Nobody owns it once the champion moves role |
The third row is the killer, and it is not specific to manufacturing. Sull, Homkes and Sull found in Harvard Business Review in 2015 that the commitments which fail most are those requiring cooperation between units, because each party assumes one of the others is carrying it. A rollout across plants is that failure mode in its purest form.
The two questions to settle before the pilot
What number moves, and what is it today?
Most pilots are judged on whether the technology worked. That is the wrong test and it produces the characteristic outcome: a pilot that is technically successful, commercially unassessed, and extended for another quarter rather than decided on. Write down the baseline before you start or you will not be able to answer the question later.
Who owns it at the second site?
Name them before the first site is finished. A rollout handed over after a successful pilot inherits none of the champion's energy and all of its assumptions.
The unglamorous part that decides it
Scaling an AI initiative across a manufacturing business is an execution problem with a technology component, not the reverse. The mechanics are the same as any cross-site commitment: one named owner, a measure that can fall, a review short enough to catch a stall in week two rather than week eleven, and an explicit owner for each handoff between functions.
That last one is the highest-return and least practised. When operations owes IT a specification and IT owes the plant a deployment, somebody owns each end and nobody owns the moment in between, which is exactly where three weeks go missing. More on that in what changes about strategy execution above 200 people.
Where Empiraa GPS fits
Empiraa GPS is a strategy and accountability platform for established organisations whose plan crosses several teams or sites, which describes most manufacturers attempting a rollout. One named owner per commitment, progress collected from where the work happens rather than chased between plants, and a fortnightly rhythm that makes a stalled site visible while there is still time to act.
It will not tell you whether to do predictive maintenance. It will tell you, in week three rather than quarter end, that site two has not started.
Related reading: what changes about strategy execution above 200 people, and why strategy execution fails.
