By Peter Sorowka, Co-CEO, Cybus

How to bring rapid AI innovation to the shop floor without putting production continuity at risk

Would you let Claude control your shop floor? Most manufacturing leaders would say no. And they would be right; no AI system should have unrestricted access to a production line, a robot, or a PLC, since there is a risk of monetary, physical, or even human damage.

Yet AI enables radically fast innovation cycles. A new application, workflow, or agent is only a prompt away. The factory, however, cannot change at prompt speed. Every new connection, configuration and application can affect quality, safety and availability.

In many plants, changes are therefore bundled into carefully planned maintenance windows, sometimes only a few times a year. As a result, valuable innovations barely reach live production because introducing them appears too risky.

So the question is not whether AI can create the next use case. It is whether your production architecture can absorb it safely.

The model is rarely the bottleneck 

Physical AI closes the loop between digital intelligence and the physical world. It perceives live conditions, makes decisions, and can influence real processes through robots, machines, or autonomous systems.

That demands much more than sending machine data to an AI model. The right data must reach the right system at the right time, with enough context to make it meaningful. A temperature value alone is not useful if the model does not know which machine produced it, which operating state the machine was in, or whether the value is current and trustworthy.

The return path is even more critical. If an AI system can initiate an action, manufacturers need to know exactly which application is acting, what it is allowed to change, and how to trace, interrupt, or reverse the resulting action.

This is why Physical AI readiness starts with the industrial data foundation, not with the model or robot.

Integration connects one application. A foundation prepares for the next.

Traditional integration is usually built use case by use case. A traceability database needs process data, so a point-to-point connection is created. A maintenance solution needs similar data, so another connection follows. The first pilot works, but every new application adds complexity and another dependency on the production environment.

A Physical AI-ready foundation takes a different approach. It decouples machines from applications, standardizes and contextualizes data once, and makes it available through a governed layer. New applications can use existing data without every team rebuilding access from scratch.

This changes the economics of innovation. The data foundation no longer supports a single use case. It supports a pipeline of future applications with less effort and lower operational risk each time.

Further, this approach keeps critical decisions close to the process. Agentic manufacturing cannot depend entirely on a permanent cloud connection. Time-sensitive and safety-relevant processing belongs at the edge while the cloud can support model development, fleet-wide analytics and orchestration. In most factories, the practical answer will be hybrid.

But the right architecture alone is not enough. Manufacturers also need a safer way to manage change.

Bring DevOps to the shop floor on OT’s terms

Software teams have learned that frequent change becomes safer when every change is tested, versioned, monitored, reversible, and small. Manufacturing needs the same capabilities, adapted to operational reality.

Configurations and data flows, managed as code, offer exactly this capability. Changes can be reviewed and validated before deployment, introduced in controlled stages, and monitored in operation. Teams need a defined rollback path if a deployment behaves unexpectedly. IT and OT must share responsibility instead of handing projects back and forth across organizational boundaries.

This is not about moving fast and breaking things. On the shop floor, breaking things can stop production or create a safety risk. Industrial DevOps is about making change controlled enough to happen continuously.

When security becomes safety

Once AI can affect physical processes, cybersecurity and operational safety converge. Every application, device, and service needs a verified digital identity. Read and write permissions must follow the principle of least privilege. Machines and controllers should never be exposed directly to AI or cloud services.

Manufacturers also need complete data lineage: which data informed a decision, where it came from, and what quality indicators and timestamp it carried. For safety-critical actions, human oversight, hard guardrails and a safe fallback state remain essential.

These capabilities must be built into the infrastructure rather than reinvented for every new use case. Otherwise, security, governance and traceability become barriers to scale.

The real measure of AI readiness

Before discussing the next model, manufacturing leaders should ask four more practical questions:

• Can a new application access live, contextualized machine data without another point-to-point project?
• Is it clear who may read data, change it, or write instructions back to equipment?
• Can a use case move to the next machine, line, or site without being rebuilt?
• Can every change be tested, monitored, and rolled back without waiting for the next production shutdown?

If the answer is no, the constraint is not the AI strategy. It is the architecture beneath it.

The factories that lead the Physical AI era will not be those with the most impressive pilots. They will be those that can turn rapid AI innovation into safe, repeatable improvements in live production without asking the line to stand still.