Industrial AI: moving pilots into plant operations

Industrial monitoring and manufacturing operations

From an experiment to a plant responsibility

Industrial AI reaches plant managers through maintenance alerts, production information and new responsibilities for scaling applications. Dustin Katje, a senior plant manager at Pella, describes permanent teams responsible for new applications, data and insights, and deployment across the manufacturing network. Jared Pfeiffer, an Augury principal client partner and former plant manager, links a credible business case to existing measures such as downtime, throughput, yield, scrap and maintenance cost. Audra Kirkland, Hexion’s vice president of manufacturing and IT, is concerned about technology providers entering chemicals without understanding its safety requirements. Protex AI co-founder Ciarán O’Mara emphasises the work of integrating tools with existing operations and documenting fragmented data. Katje’s teams start with where employees lose time, where complexity creates mistakes and where better information could improve customer service.

Pella had already described several of those applications on Oct 15, 2025. Digital assistants and chatbots provide troubleshooting support and preserve employee knowledge. Voice-guided parts picking helps workers select components for the product being assembled; the company attributes reduced training time and improved retention to that system. Cameras and software assess products, while employees receive information intended to help them learn complex processes. Pella’s Innovation Incubator contains robots, laser cutters, computer-controlled machines and 3D printers used by engineers and product developers. Vice president and chief information officer Travis Turnbull presented these investments as tools for employees serving customers. The window-and-door manufacturer, headquartered in Pella, Iowa, reported more than 10,000 employees and 20 manufacturing locations at that date. Pella did not quantify the changes in training time or retention in that announcement.

Pella company logo

Adoption figures have a population and a date

Deloitte’s May 2025 smart-manufacturing survey covered 600 executives at large companies with headquarters or operations in the United States. Fieldwork took place in August and September 2024. The companies had annual revenue of at least $500 million and more than 1,000 employees. Among respondents, 29% reported using AI or machine learning at facility or network level, while 24% reported generative AI at that scale. Pilot use was reported by 23% for AI or machine learning and 38% for generative AI. These figures describe different technology categories and deployment stages in that sample. Fifty-two percent had established a central team or working group for researching, developing and deploying smart-manufacturing initiatives. Operations leaders owned the initiative in 51% of responses; technology owners led it in 38%. Fifty-four percent reported a data standard based on a unified model, 45% used an architecture standard and 48% had a standard for training and adoption. Respondents also identified human capital and maintenance among the less mature functions.

Rockwell Automation’s May 19, 2026 research used a different sample: 1,560 management-to-executive respondents across 17 manufacturing countries, surveyed by Sapio Research. Company revenue ranged from $100 million to more than $30 billion. Fifty-nine percent reported active use of smart-manufacturing technologies, and 18% remained in pilot mode. Respondents described 34% of operations as AI-augmented and expected AI support to cover more than half by 2030. That expectation is a future deployment target, while the active-use share concerns organisations adopting the broader smart-manufacturing category. Rockwell also reported that only 43% of collected data was being used effectively and that 46% of respondents had experienced at least one cyber incident during the preceding year. Chairman and chief executive Blake Moret associated results with connecting technology, people and processes. The release identifies quality, cost, operational risk and equipment effectiveness as priorities for transformation spending.

Chemistry, equipment and quality information

Hexion’s acquisition of Smartech brought an industrial-AI business into an adhesives and performance-materials supplier. The acquisition announcement kept financial terms undisclosed and said Smartech would continue operating as an independent business serving its customers. Michael Lefenfeld, Hexion’s president and chief executive, linked the combination to the company’s resin and wood-processing expertise. Smartech chief executive Hanoch Magid described broader distribution of its technology through the partnership. The acquired systems included SmartWAX, SmartPF and SmartMDI, combining process controls with material application. Hexion described SmartMDI as a dynamic chemistry-control system for more precise resin application in engineered wood. Its cost and environmental benefits were company claims, rather than a common result established for every mill. Hexion placed the transaction within its transformation under American Securities, which took ownership in 2022.

SmartQuality added another application on May 26, 2025, when Hexion and Smartech launched it at LIGNA in Hannover, Germany. The product gives operators a real-time view of board quality and is intended to support decisions about production speed and consistency. Hexion said its models continue learning from boards produced on an individual customer’s line. The company associated the application with a possible increase in line speed of up to 3%, alongside reduced raw-material use. That figure was a stated product benefit with an upper limit, rather than a measured improvement reported for a named installation. Lefenfeld described the approach as combining data intelligence and materials science. The dated launch release also identifies Smartech as acquired in 2024, keeping the acquisition event separate from the subsequent product introduction.

Kadant Carmanah Design extended the equipment connection through a partnership announced on May 28, 2025. The Surrey, British Columbia company combined its stranding-equipment expertise with Hexion’s Smartech technology for oriented strand board production. President Todd Macey linked the collaboration to the SmartRING strander and Argus connected-technologies platform. The strander converts logs into strands before the material passes through drying, blending, forming and pressing. Argus supplies trends for log throughput, production uptime and strand quality. Lefenfeld expected wider process visibility to help stabilise strand quality and reduce variability downstream. Those improvements remained the partners’ expectations at announcement. The equipment supplier and chemistry specialist were connecting information across successive production stages, with the intended benefits extending beyond a single machine’s output. No installation-wide saving or completed fleet rollout was quantified in the announcement.

An alert becomes a maintenance job

PepsiCo’s Sacramento operation provides a maintenance example in Augury’s 2025 Spotlight case-study collection. The site monitored 50 critical assets after an earlier monitoring system had disappointed the team. Augury describes mechanics using vibration alerts to order parts and arrange repairs during planned downtime. Experienced technicians helped newer colleagues interpret the information, and support staff assisted with diagnosis by telephone. The vendor’s case study reports more than $6,000 in avoided costs, 10 hours of downtime eliminated and an 83% team response rate. It does not specify a common annual reporting period for those totals. Jakia Tadema, identified as a DVC senior resource, described a shift toward anticipating problems. Supply-chain associate Jessica Lopez associated the tool with useful preventive work. These are a vendor-selected award case and participant accounts; the figures apply to that documented site example.

Computer vision and the work behind a model

Protex AI’s June 16, 2023 account concerns a proof of concept with DHL Global Forwarding in Denmark. The system used computer vision to identify events relevant to workplace safety, including near misses and departures from procedures. Described use cases included vehicle speed, designated walkways, protective equipment and obstructed routes. Reports gave safety teams information for deciding where to change conditions or practices. DHL’s objective was proactive risk identification, supported by continuous site visibility. The partnership announcement did not provide an independently verified accident-reduction rate for the pilot. Protex also made privacy and regulatory-compliance claims; those remain the provider’s assertions. The case concerns information supplied to safety personnel, with the operational response sitting in the organisation using it. It is distinct from machine-condition monitoring, even though both applications turn incoming data into events for staff to investigate.

Voxel51’s July 15, 2025 Protex AI case study follows the development work supporting those visual systems. Protex processed video from more than 1,000 cameras across over 100 customer sites. Its team moved from scripts and fragmented tools toward centralised data inspection, curation and model analysis. Raw footage was checked for corruption or network problems before training subsets were selected. Engineers then compared model performance with the underlying images, including true and false detections. Around 10 internal plugins supported tasks such as filtering data, passing annotations between tools and running inference jobs. Patrick Rowsome, head of computer-vision operations, valued the ability to customise those workflows. Voxel51 reported five-times-faster model iteration. That is a claimed change in the development process, with a different denominator from workplace incident rates or factory throughput.

Engineering assistance and deployment scope

Siemens’ Oct 24, 2024 Industrial Copilot announcement distinguishes a product’s existing users from its potential audience. More than 100 companies, including Schaeffler and thyssenkrupp Automation Engineering, were using the assistant, Siemens said. A separate 120,000-user figure referred to engineers already using the TIA Portal engineering environment who could adopt the generative-AI tool. thyssenkrupp Automation Engineering planned a global rollout beginning in early 2025; the announcement did not establish that later rollout as completed. The battery-inspection example combined sensors, cameras and measurements with assistance for documentation, data management and sensor configuration. Siemens associated the copilot with reducing repetitive engineering work. Its claims about creating panel visualisations in 30 seconds and code requiring 20% adaptation describe stated capabilities, rather than a universal result for all machines. Microsoft supplied Azure OpenAI Service within the collaboration.

NIST researcher Michael Sharp addresses data suitability directly in a Feb 4, 2025 industrial-AI article. His assessment covers the information, assumptions, rules and shortcuts involved in training, testing, development and deployment. He identifies simplified simulations, selected ideal examples and excessive smoothing as ways a dataset can lose contact with actual manufacturing variation. Different machines, shifts and departments may contain relevant information that a single dataset misses. Rare events, including breakdowns, can also be underrepresented. Sharp recommends identifying such gaps and considering how the system will be monitored when those conditions occur. His emphasis is on relevant, representative data, rather than dataset size alone. The guidance also considers missing observations, inadequate variation and gaps between operating states. It is an assessment approach for industrial users, without a claim that any named manufacturer has already met those conditions.

Evaluation in a controlled manufacturing environment

NIST’s Collaborative Robotic Operations Workcell supplies a physical setting for developing and evaluating industrial monitoring. In its May 2024 description, the Industrial Artificial Intelligence Metrology and Management team worked with cybersecurity researchers to adapt the benchtop workcell as a source of manufacturing data. CROW has multiple stages and 10 major physical components, including collaborative robots, conveyors, inspection cameras, sensors and digital logging equipment. Robotic arms coordinate production and evaluation of cyclic product streams. The research spans digital communications, product quality and human interactions, with anomaly detection and process-error prevention among the intended applications. NIST planned open access to the data so developers and manufacturers could build and test domain-appropriate tools.

The NIST AI Risk Management Framework, released on Jan 26, 2023, addresses organisational responsibilities across AI development and use. It is voluntary guidance for organisations designing, developing, deploying or using systems. NIST describes AI as socio-technical: outcomes depend on human behaviour and organisational context as well as technical components. Training data can change, sometimes unexpectedly, altering system behaviour. The framework separates a discussion of risk and trustworthy characteristics from a core organised around four functions: govern, map, measure and manage. These functions can be applied to particular use cases and stages of the AI life cycle. A companion playbook suggests approaches to using the framework. NIST developed it with public and private participants, drawing on about 400 sets of formal comments from more than 240 organisations.

Generative systems add their own evaluation questions

NIST’s July 2024 Generative Artificial Intelligence Profile extends that framework for generative applications across sectors. It describes profiles as adapting the framework’s functions to a setting or technology, taking account of the user’s requirements, risk tolerance and resources. Its introduction identifies governance, content provenance, pre-deployment testing and incident disclosure as the working group’s four main considerations. Risk can arise during development, deployment, operation or decommissioning, and can occur at model, application or wider ecosystem level. The profile distinguishes technical factors from human behaviour and interactions with systems. One identified risk is confabulation: confidently stated content that is erroneous or false. The document treats risk as involving both the likelihood of an event and the magnitude of its consequences. The profile also discusses additional human review, tracking, documentation and management oversight for generative systems, together with consideration of third-party models and services.