The World Economic Forum runs an annual scorecard for manufacturing called the Global Lighthouse Network. Companies apply, get judged by an independent panel, and if their factory truly stands out on productivity, sustainability, or use of AI, they get named a "Lighthouse." By early 2026 the network had grown to 223 sites across more than 30 countries and 40 industries.
That sounds like a lot of proof that AI works in factories. It is. But a new WEF analysis points to an uncomfortable detail hiding inside that success story: of all those award winning companies, only 23 managed to take the improvements from their best factory and successfully copy them into three or more other plants. The rest have one shining factory and a lot of ordinary ones sitting right next to it.
This is not a small technical hiccup. Manufacturing consultants have described this exact trap for years under the name "pilot purgatory," where a company runs an impressive test project, gets praised for it, and then never manages to repeat it anywhere else in the business.
What makes this moment different is the speed AI is moving at. Inside the current crop of Lighthouse factories, ordinary data driven AI now shows up in roughly six out of ten of their top projects, and generative AI, the kind that writes and reasons, jumped from about 9 percent of projects in 2024 to 23 percent in 2025. The best factories are pulling ahead faster than they used to. If the average factory cannot copy what the leaders are doing, the gap between the two groups only gets wider.
The scale of the problem becomes clearer once you zoom out. Industry estimates put the number of factories worldwide at roughly 7.5 million. Two hundred and twenty three Lighthouses is a rounding error against that number, which means the lessons from these standout sites have barely begun to reach the broader manufacturing world.
The WEF's suggested fix is what it calls a "floor and frontier" approach. The "floor" is a shared baseline every factory in a company's network gets: the same way of measuring performance, the same basic digital skills for workers, and clear rules for who owns which decision. The "frontier" is a smaller set of sites allowed to test the newest and riskiest ideas. The floor is what makes copying possible later; the frontier is where new tricks get invented first.
A real example makes this concrete. ACG Packaging Materials, a pharmaceutical packaging plant in Shirwal, India, took equipment that was 60 years old and fed it two years of quality data through machine learning models. The result was a 37 percent jump in first pass yield, meaning far fewer batches needed rework, and a 43 percent cut in the time needed to switch equipment between product runs. None of that required new machines. It required organized data and a clearly defined problem.
Mettler Toledo's site in Changzhou, China took a different path, pulling in universities and suppliers alongside its own staff, then successfully extended what it learned to six more locations. That is the pattern winners share: they build the copying mechanism into the project from day one, rather than trying to bolt it on after the fact.
For any company running more than one site, the message is blunt. A single impressive AI project proves that AI can work. It proves nothing about whether your organization can make it work everywhere else. The real advantage now belongs to whoever builds the muscle to repeat success, not just produce it once.