Manufacturers continued to face rising costs and supply chain pressures in the third quarter of 2026. According to the National Association of Manufacturers’ Q3 Manufacturers’ Outlook Survey, increased raw material costs remained the top business challenge for the second consecutive quarter, with rising healthcare costs and trade uncertainty rounding out the top three. Respondents expect raw material and other input costs to increase 5.0% over the next 12 months. Among manufacturers surveyed about the conflict in the Middle East, 60.6% said conditions had not improved from the previous quarter, while 33.2% said challenges had worsened.

The pressure extends beyond any single market. A European Commission survey of more than 12,000 manufacturing enterprises found that firms with trade or production activities outside the EU were adjusting inventories, sourcing, prices and contracts in response to a more volatile and fragmented trade environment.

The Commission identifies geopolitical realignment, dependencies on China, US trade uncertainty and disruption to shipping routes in the Middle East as key features of that environment.

These conditions increase the value of faster insight and more adaptive operations. Yet the central AI challenge has shifted toward scale. Deloitte’s 2026 manufacturing research found that 84% of respondents were generating measurable value from AI, while only 20% of use cases had been scaled consistently across sites or the enterprise.

A successful pilot can improve one line or one plant, but enterprise value depends on whether that capability can be replicated across different assets, systems, regions and operating conditions without rebuilding it each time.

Why successful pilots struggle to travel

The difficulty often begins with the way a pilot is designed.

Under pressure to demonstrate an early result, a team may connect an AI model directly to local supervisory control and data acquisition systems, programmable logic controllers or other plant technologies. The use case performs well in that environment because its asset names, data pipelines and interfaces have been configured around the specific plant.

Problems emerge when the organization attempts to move it elsewhere. The same motor, temperature signal or vibration measure may have a different name at another facility. Equipment may come from another supplier, operate through a different control system or sit within a locally adapted production process.

Three design choices commonly limit scalability:

• Local hardcoding: Asset names, data mappings and interfaces are built around one facility rather than a reusable data model
• Plant-specific infrastructure: The pilot runs on dedicated local hardware without considering how compute, licensing and support costs will change across dozens of sites
• Standalone insight: The model produces an alert or dashboard but remains disconnected from the maintenance, production, inventory and procurement workflows needed to act on it

Each choice may accelerate the pilot, but together they create additional work when the organization attempts to scale. The result is often a technically successful use case with an unsustainable enterprise cost.

Manufacturers need to consider replicability and total cost of ownership during the initial design, rather than after the first plant has gone live.

Create a common foundation without ignoring plant differences

A scalable approach begins with a digital blueprint covering common data models, edge-to-cloud architecture, deployment patterns, governance and operational integration.

A Unified Namespace (UNS) can provide a consistent way to organize, name and contextualize operational data across facilities. Using a publish-subscribe architecture, it can create a common real-time view of data that different plant and enterprise systems can access. This reduces the need for repeated point-to-point integrations and can provide a more scalable data foundation for advanced AI.

The architecture must also distinguish between centralized and edge processing. Some workloads need to run locally, while other capabilities can be managed centrally. Time-sensitive processing may need to remain close to production at the edge, while model management, cross-site analytics and enterprise governance can sit in cloud or centralized platforms. Containerized deployment can then make those capabilities easier to move, update and manage across different environments.

Technology alone will not resolve the challenge. A scalable operating model also needs:

• A central AI function that owns the architecture, governance standards, reusable components and enterprise performance measures 
• Plant teams with the authority to adapt solutions to local operating conditions 
• Integration with manufacturing execution, maintenance and enterprise systems 
• Clear ownership of adoption, process change and operational outcomes 

Temperature, humidity, input material quality, regulation and operating practices can all affect machine behavior. A UNS can provide context for operational data and support more consistent deployment across plants. Its hub-and-spoke architecture can reduce the need to build new point-to-point integrations for each system. Effective scale requires a reusable core combined with controlled localization, rather than a rigid copy-and-paste model. 

Trust depends on the reality of the plant

Data can become an afterthought when teams are under pressure to demonstrate progress quickly. Some pilots may rely on synthetic or sample data, creating challenges when the solution moves into a live plant environment. Enterprises are taking a more deliberate approach to data governance and cleansing, but those steps alone may not capture the full operational context.

Industrial data often provides only a partial view of the physical environment.

A maintenance system may not show that a bearing or seal was recently replaced. Wiring diagrams may be outdated. Experienced operators may know that a machine behaves differently under a particular load, but that knowledge may never have been recorded. These cases can be overlooked in data-led initiatives, creating an incomplete or inaccurate picture of conditions on the plant floor.

When incomplete data causes a model to recommend the wrong course of action, trust can deteriorate quickly. That risk is heightened by sensor drift, model drift, fragmented systems and unreliable agent behavior, making a solution that loses the confidence of operators in one plant difficult to scale across the enterprise.

This makes plant personnel an essential part of AI initiatives. Operators should help validate the data, test recommendations and identify the undocumented context that affects performance. Their feedback can improve the model while plant-level champions help colleagues understand how the capability supports their work.

The process also must account for operational pressure. Employees in continuous manufacturing environments cannot simply stop a furnace, compressor or production line to accommodate a technology rollout. Adoption plans must reflect the limited bandwidth of teams whose primary responsibility is to keep production running.

Move from insight to controlled action

Predictive maintenance, computer vision and production optimization remain important, but industrial AI is beginning to support more coordinated decisions that close the gap between insight and action.

Digital twins can help teams evaluate production changes before applying them to live operations. This can be valuable when raw material shortages, energy costs or demand changes require a manufacturer to rebalance lines, alter throughput or adjust process parameters.

Agentic AI can coordinate the steps involved in responding. An agent might interpret a quality or equipment signal, retrieve maintenance records, check production priorities and parts availability, then prepare an action plan. Depending on the risk, it could route that plan for approval or initiate permitted actions through connected systems.

Closing the gap between insight and action is important, but enterprises should proceed carefully as evidence builds and confidence grows. A practical framework has three levels.

1. Autonomous execution within defined limits

   Low-risk and reversible actions, such as routine inventory replenishment or warehouse heating and cooling adjustments, may be suitable for autonomous execution once the system has demonstrated consistent performance.

2. Human approval before execution

   Adjusting assembly-line speed, process temperatures or chemical feed rates can affect throughput, scrap, equipment effectiveness and product quality. AI can develop the action plan, but an operator should approve it before changes reach the control environment.

3. Human-controlled execution

   High-cost, safety-critical or difficult-to-reverse actions require accountable human control. Restarting a plant after a blackout, changing a core formulation or overriding a safety control may be supported by simulations and recommendations, but execution should remain manual.

   The expected benefit should always be considered alongside the cost of failure. A high level of autonomy may be technically achievable, but difficult to justify when one incorrect action could damage equipment, waste a batch or create a safety incident.

Governance must be inclusive

Industrial AI governance needs to cover the entire decision chain: data, models, agents, integrations, approvals and outcomes.

For each use case, manufacturers should define:

• Who owns the capability and its business result 
• What data the model or agent can access 
• Which systems it can update or control 
• When approval or escalation is required 
• How activity will be logged and audited 
• What happens when performance declines or an incorrect action occurs 
• How model tolerances, hallucinations and drift will be monitored as operating conditions change
• Operational technology security must be built into the same governance model because scaling AI can create new connections into protected plant environments. Architecture teams need to preserve network segmentation, tightly control access and ensure AI services do not create unmanaged routes into production systems.

Measure whether AI can scale

There are multiple tools, approaches and use cases, so leaders need to focus on those that are relevant, scalable and tied to business problems. My recommendation is to break those problems into smaller components that can be addressed through pilots and measured for business value. Some of the areas we are working on across the industry include:

• Overall equipment effectiveness 
• First-pass yield 
• Scrap and waste 
• Changeover performance 
• Conversion cost per unit or ton 
• Energy consumption 
• Cutting latency
• Safety 

We must keep in mind that while a pilot may be judged by the improvement it delivers in one plant, an enterprise program needs additional measures that show whether the capability can travel.

Leaders should also track delivery and scalability indicators, including time to first insight, pilot-to-production velocity, stabilization time, payback and the total cost of operating the capability across the estate.

Replicability itself should become measurable. How much customization is required to apply the data model at another plant? How much legacy integration work must be repeated? Does the deployment architecture support central updates? How do regional conditions affect model performance?

AI error rates and the potential cost of an incorrect decision should also determine the level of human oversight. These measures expose the difference between a strong demonstration and an enterprise capability. 

Design for the enterprise from the beginning

The next phase of industrial AI will depend on connecting plant-level intelligence with repeatable architecture, governed action and measurable business performance.

Manufacturers that address data consistency, edge-to-cloud deployment, workflow integration, change management and security from the outset will be better positioned to extend successful use cases across their networks. The same foundations will support digital twins, intelligent agents and progressively more autonomous operations.

The strongest programs combine enterprise standards with the experience of the people operating each facility. That balance allows manufacturers to scale faster without losing sight of the local realities that determine whether AI works in production.