Manufacturers face significant challenges expanding AI across multiple facilities, where inconsistent data and systems can complicate deployment.
By Rob Williams, Centric Consulting
Only 23 of the 223 most advanced manufacturing sites in the world have managed to expand their AI and digital gains to three or more locations, according to recent data from the World Economic Forum, a sign of just how rare successful scaling is industry wide.
A pilot that works at one plant routinely fails to translate to the next, even for enterprises best equipped to make it work. The AI usually works as expected, but differences in the second site’s data and systems make the pilot difficult to replicate.
On a manufacturing AI rollout, momentum stalls when nobody plans for how different the second site is from the first.
Many manufacturing AI pilots follow a similar path. An IT team responsible for building out the company’s data infrastructure brings operational data into the fold, picks a tool, and chooses a use case that will deliver visible value on the floor. Budget gets approved; the pilot plant delivers results, and everyone involved treats those early wins as proof the company has built something repeatable.
The reason it works has less to do with technology than scope. A pilot is typically built around one site’s specific data, cleaned and organized by hand to prove the concept, not a broad, reusable data foundation. That narrow scope is exactly what makes a first pilot achievable without the enterprise-wide infrastructure most manufacturers haven’t built yet.
AI relies on each plant’s equipment, data labels, and systems of record. Those inputs naturally vary by site, often far more than teams anticipate, so the same AI solution is rarely as replicable as it first appears.
In one manufacturing AI rollout our team worked on, a different source for local operational data and a non-standardized file structure turned what should have been a simple replication into weeks of extra customization, roughly doubling the rollout timeline at the second site.
This kind of variation shows up across manufacturing pilots once they move past a single site, and it’s a reality companies need to account for as they build a unified data model for the enterprise.
What looked like a repeatable win at the first site becomes a custom project at the next. A budget built around “scale this everywhere” quietly becomes a budget for “make this work at one more place,” and momentum drains out of the initiative long before anyone officially calls it stalled.
The data foundation required to scale AI often doesn’t exist yet, even after a successful pilot.
Recent data shows manufacturers lag well behind other mid-market companies on basic data infrastructure: just 27 percent have a data warehouse or data lake, versus 60 percent of mid-market companies overall, and none of the manufacturers surveyed had yet adopted machine learning platforms.
The second site’s data rarely looks the same as the first, and without a shared model to translate between them, every new location becomes its own project.
Scaling works when the data model comes first and the pilot is treated as one use case built on top of it. That means building the underlying architecture broadly enough from the outset to support use cases and sites the company hasn’t identified yet, rather than optimizing narrowly for the first win.
It also means budgeting for a second site as its own discovery phase, with its own timeline, rather than assuming it will move at the speed of copy-paste. Planning for that variation upfront, standardizing what can be standardized while leaving room for what has to stay local, avoids the quiet cost and schedule overruns that kill momentum after a strong first result.
Pilots are typically built around one site’s specific data and systems. A second site’s OT network, data sources, and file structures are often different enough that replicating the pilot requires nearly as much custom work as building it did the first time.
There’s no universal answer, but it should be a deliberate choice. Without that decision made upfront, the answer tends to get discovered expensively, one site at a time, as each rollout takes longer than expected.
Treat the second site as its own discovery phase rather than a replication. Budgeting time and cost for data mapping, integration, and customization at each new site, rather than assuming a fixed, repeatable cost, prevents the timeline surprises that stall momentum.
It means building the architecture that connects operational data to business systems broadly enough, from the start, that it can support additional use cases and sites as the organization learns what works, rather than building something narrowly scoped to one pilot’s needs.
Fixing stalled scaling means treating data infrastructure as an enterprise investment from day one, not a one-off pilot asset. It’s the difference between a single win and joining the roughly one in ten advanced manufacturing sites that have managed to expand it beyond two locations.
About the Author:
Rob Williams is the supply chain lead at Centric Consulting, an international business and technology consulting firm. He brings more than 20 years of experience in manufacturing operations and supply chain, from frontline operations through executive management, and helps manufacturers improve performance and streamline operations. He is passionate about balancing technology with the physical realities of manufacturing and the people who make it run.
Read more from the author:
Transform Your Factory Floor with Automation in Manufacturing | Centric Consulting, November 2025
6-Step Guide to Implementing Value Stream Mapping in Manufacturing | Centric Consulting, April 2026
