AI after the pilot phase

The most successful AI pilot in a factory may also be the least useful guide to what happens next. It is usually built around a clearly defined problem, supported by an enthusiastic local team and protected from much of the inconsistency across the wider business. The data has been curated, and the team wants the project to work.

Scaling removes that protection. The system must operate across plants with different equipment, data standards and local practices, while fitting into live production without disrupting safety, quality or delivery. The question is no longer whether the technology can produce an accurate result, but whether the manufacturer can make that result part of the way the organization operates.

This is why the gap between proof of concept and enterprise deployment is proving so difficult to close. Pilots test capability, but scale tests ownership, governance and operational discipline. Treating the second phase as a larger version of the first soon reveals that technical success was only the beginning.

A pilot proves less than it appears

The protected conditions that make a pilot achievable can conceal the weaknesses that later prevent it from scaling. One plant may have relatively clean data and a strong regional champion, while another depends on fragmented systems and local workarounds. The model can be identical and still encounter a completely different operating environment.

Maggie Slowik, Global Industry Director for Manufacturing at IFS, sees that difference as one of the main reasons early successes fails to travel. “You might have one site with relatively clean operational data, strong leadership support and people driving the initiative locally, but then you try to roll that same model out across multiple factories and the conditions are completely different,” she explains. “Sometimes you have a regional champion who drives success in one location, but you do not have the same buy-in or operational readiness elsewhere. Without that broader governance structure and rollout strategy, scaling those AI initiatives becomes very difficult.”

Manufacturers therefore need to design pilots with scale in mind rather than treating enterprise deployment as a later concern. Other plants should be involved early enough to identify process differences and local constraints. A pilot should remain focused, but not so isolated that it proves something the wider organization cannot reproduce.

Will Dutton, Director of Supply Chain Solutions at UiPath, makes the point more bluntly. “Most pilots fail before they are ever built, because no one tied them to something the C-suite loses sleep over,” he says. “If your AI experiment is not aligned to a board-level outcome, it will never earn the investment needed to leave the lab. Scaling AI across plants is not a copy-paste exercise. When you roll out across regions, every site wants something slightly different, and the winners solve for mass customization.”

Ownership changes when AI enters operations

A proof of concept can survive with diffuse ownership because its purpose is exploratory. A digital team may run the project, a plant manager may sponsor it and a vendor may provide much of the technical support. Once the system starts influencing production planning, maintenance decisions or quality control, that arrangement is no longer sufficient.

Operational AI needs an executive owner who is accountable for the outcome rather than merely responsible for funding the technology. That owner must resolve conflicts and ensure the system continues to deliver after the project team has moved on. Dutton argues that manufacturing already has governance structures that can be extended to AI rather than replaced by something entirely new.

“The pivot from pilot to operations changes the governance question entirely,” he says. “In the pilot phase you can get away with diffuse ownership, but AI needs an executive owner accountable for outcomes, not just budget. If everyone owns it, no one does. Manufacturing already runs rigorous frameworks for safety, quality and change control, and AI should not sit outside those. Finance should also be in the room from day one, because if finance signs off on impact, it must be part of the governance structure rather than being consulted at the end.”

It also changes how value should be measured. Too many pilots select a problem that is small enough to be safe but not important enough to justify wider investment. Manufacturers need to define the business measure before deployment, establish who will validate it and decide how benefits that do not appear directly on the profit and loss statement will be treated.

Scaling exposes the real operating system

The further AI moves into production, the more it becomes entangled with the rest of the manufacturing environment. Data must pass between platforms; recommendations need to trigger workflows and decisions made in one part of the operation can create consequences elsewhere. A contained use case begins to reveal a network of dependencies.

Phil Hadfield, UK Managing Director at Rockwell Automation, describes this as the natural consequence of digital maturity rather than evidence that transformation has gone wrong. “As digital systems become more deeply integrated into manufacturing environments, the number of interactions between them increases significantly,” he explains. “Each connection between platforms, assets and data sources introduces new dependencies and new considerations around performance, resilience and control. That complexity is not necessarily a sign of failure, but it does require a different level of operational discipline because manufacturers are managing highly interconnected environments rather than isolated technology deployments.”

This is where legacy systems and inconsistent data structures become impossible to ignore. A pilot can work around missing information or rely on a local expert to interpret an output. Enterprise deployment needs a repeatable way to access, contextualize and govern data across the organization. Without that foundation, usefulness deteriorates from site to site.

Slowik argues that AI is forcing manufacturers to confront problems they have postponed for years. “Companies have spent years collecting operational data and deploying sensors, but many still do not have a clear strategy for how to use that information effectively. AI is creating pressure because organizations understand that these systems are only going to be valuable if the operational data underneath them is reliable and connected. That is pushing manufacturers to revisit how they manage operational visibility, data integration and digital strategy across the business.”

Trust must survive the factory floor

Technical performance is only part of operational acceptance. Employees need to understand how the system reached a recommendation, especially when it contradicts experience built over years on the line. If the explanation is weak or the data cannot be traced, operators are likely to return to the spreadsheets and manual methods they already trust.

That resistance is often misread as reluctance to adopt technology. Production teams are responsible for maintaining continuity and are cautious about changes that could destabilize the process. Trust develops when people can challenge the output and understand when human judgment still takes precedence.

Slowik sees domain expertise becoming more important as AI becomes more capable. “In manufacturing, the challenge is rarely just about having a more capable model. It is about understanding the operational context in which that model is applied. You need to spend time with the people running production, planning and supply chains to understand where decisions are made, where variability exists and where bottlenecks really sit. Generic AI models without that domain understanding are not going to solve real operational manufacturing problems.”

Dutton adds that systems should absorb local exceptions rather than treating them as errors. Operators often know which workarounds are essential, and that knowledge must become part of the design. A scalable system is not one that eliminates every variation, but one that knows which variations matter.

From deployment to operational capability

Manufacturers are entering a more disciplined phase of AI adoption. The technologies are increasingly proven, but the measure of success is shifting from deployment to consistent performance in live environments. Reliability, resilience and integration now matter more than the number of pilots completed.

Hadfield captures that change clearly. “The focus is less about proving that the technology works and more about ensuring that these systems can operate consistently inside real manufacturing environments,” he adds. “Reliability, resilience and integration are becoming the measures of success rather than deployment itself. Organizations are increasingly measuring success not by whether individual technologies have been adopted, but by how effectively those technologies integrate and perform within the broader operational system.”

The manufacturers moving beyond pilot purgatory are not simply running more experiments. They are assigning ownership, involving finance before the benefits are claimed and building AI into the governance structures that already protect production. They are also accepting that scale requires adaptation because the system must work with the reality of each plant rather than the assumptions of the first successful trial.

A pilot can show that AI can produce an answer. Enterprise deployment begins when the organization can trust that answer, act on it consistently and remain accountable for the result. That is the point at which AI stops being an innovation project and becomes part of manufacturing operations.