Capital is flowing into industrial AI faster than into almost any other category of plant investment. About 70% of manufacturers have an active AI initiative underway. And it's for good reason: Predictive maintenance models are spotting bearing failures before failure can occur, computer vision systems are catching defects that human inspections miss, and intelligent automation is rebalancing production schedules in real time.

So why don’t the results match the spending?

McKinsey research found that only one-third of organizations have moved AI past the pilot phase into production. Manufacturers keep running into the same paradox. The demos work, the proof of concepts and business cases are impressive, but the second a successful pilot tries to scale across plants and workflows, it stalls out.

What most leaders haven’t fully realized is that the problem isn’t the AI itself. In fact, the technology is more than capable of delivering. The real issue is that most manufacturers are trying to scale that AI on top of fragmented data, disconnected systems and operational processes that were never designed to support it.

The great divide

McKinsey calls this paradox the “great divide” between AI investments and ERP systems that actually run the business, and their findings should land especially hard for industrial leaders.

Almost half of IT organizations are shifting budget into generative AI while reducing investment in the core infrastructure that those AI systems depend on. Only about 40% of companies report any enterprise-level EBIT impact from AI, despite roughly 80% using gen AI in at least one function. McKinsey calls the result “pilot purgatory,” a state where use cases proliferate but never get the data, processes or system integration needed to scale.

This divide is even sharper for manufacturers than for the average enterprise. Industrial AI doesn’t live in a sandbox of experimentation, but downstream of ERP, MES, historian databases, SCADA systems, IoT platforms and quality management tools that were each built at a different time to serve other purposes.

Think about what that means in practice. A predictive maintenance model is only as good as the sensor data, work-order history and asset hierarchy feeding it. And a demand-forecasting agent relies on the connection between order data in ERP, production status in MES and supplier signals in procurement. When those layers don’t talk to each other, the AI doesn’t scale at all, but simply produces unreliable answers quicker than the old systems did.

Why successful pilots aren’t holding up in production

The reason a pilot succeeds, yet a rollout fails, is almost always the same. The pilot was carefully scoped around clean, available data in one plant or one line, but the scale-up is not. Three patterns show up over and over in manufacturing:

1. Data that can’t be trusted across sites. Plants have different MES versions, naming conventions and tag structures in the historian. This means a model trained on one site’s data can’t apply the same training to the next. Stitching the data together is a time-intensive process that delivery teams rarely budget for.
2. Legacy infrastructure that wasn’t built for AI workloads. Older ERP and control systems weren’t designed to expose data through APIs, support streaming or integrate with cloud-based model endpoints. Manufacturers end up bolting middleware on top of middleware, and every integration point becomes fragile.
3. Workflows that haven’t been redesigned. Throwing AI on top of existing processes without updating the workflow won’t work. Let’s say you have a defect-detection model running on the line. The AI can help spot issues faster and more accurately than before. But if those defects still flow into the same quality review process (same people and schedule), the bottleneck just moves down the line to people already stretched thin.

Operational alignment is more important than the latest tool

In order for manufacturers to succeed, AI scaling needs to be seen as an operational program first and a technology program second. Here are the recommended priorities:

1. Build the data foundations before adding more use cases. Use a common data model across plants, governed sensor and master data, lineage tracking and quality monitoring. This is because every new use case isn’t repaying the same data cleanup tax. Data readiness is the strongest predictor of whether the next pilot will scale.
2. Connect AI to the systems of record. AI agents and ERP have to be designed together, not in parallel. That approach has to extend to MES, historian and supply chain systems as well. AI that can’t read from and write to the systems running the plant won’t change how the plant runs, even with the most sophisticated model.
3. Redesign workflow around AI. Start small. Pick a contained value stream, such as unplanned downtime for a critical asset class. Then look at the people who deal with that problem today, like the operator on the line or the reliability engineer diagnosing the issue. Map out what each of them does now, where they wait on each other and where decisions get stuck. Now you can redesign that workflow around the AI. Once it works for one asset class in one plant, you have a playbook you can take to the next one.
4. Assign individual accountability. Steering committees and shared ownership tend to lead to confusion and a lack of consistency in follow-through. Name an owner for each use case, tie funding to outcomes like OEE, scrap rate and on-time delivery and review on a regular cadence.
5. Invest in the people. A model can only deliver value if the people around it actually use it. This means the human side of the rollout is as important as the technical side. Train the plant floor on basic AI literacy so operators understand what the system is recommending and why. For line leaders, teach change management and walk them through how their daily routines and decisions are going to shift. And at the executive level, model the change and use the tools yourself.

Act now or be left behind

In the next phase of manufacturing transformation, AI models will be increasingly commoditized. The key to success will be locking in on the quality of the data foundations, system integration and operational discipline to run AI reliably across every plant in the network.

If manufacturers are still treating AI as a portfolio of pilots, they will keep cycling through proofs of concept and miss out on the compound advantages of effective AI quarter after quarter: lower downtime, higher yield and faster cycle times in a sector where margins are thin and capacity is the constraint. That compounding benefit can make or break the entire game.

Fix the foundation, and AI will both scale and deliver.

Michael Simms is the Vice President of Data & AI at Columbus, and is a seasoned technical manager who has been developing data and artificial intelligence solutions for nearly three decades. He has been at the leading edge of AI, Data, ERP and other emerging technologies. He plays a principal role in architecting and implementing projects from creation through go-live. Simms also excels at creating and supporting offerings in the analytics/digital transformation space, specifically for Gen AI, machine learning, and data science. His extensive expertise includes data architecture, data migration, data engineering, and AI.