New Data Flywheels Will Determine Automation Engineering AI Winners and Losers

Manufacturing companies are being offered a plethora of Artificial Intelligence (AI) options within the market now, as industrial automation suppliers such as ABB, Emerson Electric, Rockwell Automation, Siemens, and SUPCON strive to capture new customers and position themselves as AI leaders. This race for market share offers gains in productivity for customers, but for the suppliers, capturing the most clients brings something more important than revenue: data. AI tools are only as good as the data they train on, so being able to access the most data will make the best models.

Access to operational data is the critical first step of industrial AI training. Industrial automation suppliers have access to this through their own manufacturing practices, supplying critical training data of Programmable Logic Controller (PLC) code, tribal knowledge, and product workflows. While access to these data is helpful, it only captures the manufacturing process of industrial machinery, which is only 4% of the total PLC market. (See ABI Research’s Industrial Automation Hardware market data (MD-IAH-25.)

This misses the unique aspects of other discrete industries such as automotive and aerospace & defense. Similarly, internal data completely misses process industries, which have a different set of activities entirely. Training data pitfalls reduce product usability, making AI output prone to hallucination and taking longer to verify from users, reducing the odds of tools being used.

The need to capture more data gives industrial automation suppliers different paths to capture more customer data.

Co-innovation emphasizes that customers are now partners in innovation; enabling customers to feel like they are an important part of new developments from suppliers using their data. Building In-house models carries large costs associated with training and inference. Therefore, co-innovation frameworks are best for companies that have expansive resources for in-house investment, such as Siemens, which has co-innovation partners in NVIDIA and PepsiCo and can leverage its resources to strive for the best general industrial AI model.

A different form of co-innovation is building partnership ecosystems, allowing external AI companies to enable innovation within the network. Bosch Rexroth has built such an ecosystem with ctrlX World, allowing customers to access a breadth of AI models at a lower level of investment. However, outsourcing AI brings limited control of model development, so companies pursuing this path must have stringent quality requirements for models. Leveraging a partnership network is best for firms already focused on interoperability, such as Schneider Electric.

SUPCON has taken the path of developing cooperative data networks across industry and academia with its Industrial AI Data Alliance in China, featuring over 130 partners. Siemens uses this strategy in its relationship with European machine builders and has called on the need to develop more of these alliances. Data partnerships benefit companies interested in capturing the data and progress across the spectrum. Such a model is best for companies aiming to provide a network-driven product, where system integrators are an important part of bringing value to market.

Open sourcing is the most radical option, asking customers to provide their data openly is difficult. While virtual PLC newcomer Autonomy Logic successfully built an open-source community of over 9,500, that may be due, in part, to its roots as an academic project. Open sourcing is an alternative only viable to community-driven upstarts and laggards in the AI race, such as Mitsubishi Electric and Omron, which would likely be making a smaller investment to support other models, mirrored by their slow adoption of digitalization.

Accessing data enables a new AI flywheel, with expanding relationships enabling better models. This improves competitive viability and new logo wins, which, in turn, generate new data to further improve the model.