AI サービスが一時的に利用できないため、復旧後に翻訳を補完します。
This article is sponsored by Poka and was written, edited, and published in alignment with our Emerj sponsored content guidelines. Learn more about our thought leadership and content creation services on our Emerj Media Services page. Digitized records can still leave factory operators piecing together information from disconnected systems. Differences in equipment, workflows, and worker experience across plants make that information harder to apply consistently. A 2026 paper in Data in Brief identifies inconsistent data-exchange standards as a barrier to interoperability in manufacturing. Guidance from the Australian Signals Directorate and international partners emphasizes safety and operational knowledge as foundations of operational technology security. Manufacturers introducing AI into physical workflows must therefore account for both disconnected data and the operating environment’s safety requirements. Emerj recently hosted manufacturing leaders on the Emerj AI in Business Podcast to examine how those constraints shape industrial AI. Guests included Antoine Bisson, CEO and Co-Founder at Poka; Scot Burdette, Global CIO of Measurement and Analytics at ABB; and Jeff Witt, Global Process Digital Authority at GE Vernova. Their conversations examined why industrial AI is not enterprise AI, and why it has to be built around frontline workers and the realities of physical operations. This article examines four insights for leaders evaluating industrial AI that scales: Purpose-built industrial AI for controlled physical operations: Match the system’s architecture and decision boundaries to equipment, safety, and production requirements before introducing autonomy. Connected operational data for context-specific guidance: Link relevant machine history, work instructions, training requirements, and production conditions to support informed frontline decisions. Workflow-embedded AI assistance for faster frontline problem-solving: Bring issue reporting and contextual troubleshooting into the worker’s existing task, with human review of consequential decisions. Standardized workflows for repeatable multisite deployment: Optimize recurring processes and establish shared data structures so additional plants can reuse the foundation while retaining local context. Listen to the full episodes: Episode 1: How Industrial Leaders Are Redefining AI for the Factory Floor – with Antoine Bisson of Poka Guest: Antoine Bisson, CEO and Co-Founder at Poka Expertise: AI-Native Strategy, Industrial Software, Manufacturing Technology, Software Engineering Brief Recognition: Antoine Bisson is CEO and Co-Founder of Poka, a manufacturing software company acquired by IFS in 2023. After more than a decade as CTO, he became Chief AI Officer in 2025, leading Poka’s company-wide AI transformation, before becoming CEO in 2026 with a mandate to build an AI-native organization. Earlier in his career, he worked as a Program Manager at Microsoft and held software engineering roles at Tagged and Ubisoft. He is also a Limited Partner at Inovia Capital and Blank Ventures. He holds an Honors Bachelor of Software Engineering from the University of Waterloo. Episode 2: The Real-World Conditions That Shape Industrial AI – with Scot Burdette of ABB Guest: Scot Burdette, Global CIO of Measurement and Analytics at ABB Expertise: Enterprise IT strategy, technology integration, manufacturing digital transformation, M&A integration. Brief Recognition: Scot Burdette leads technology strategy and modernization as Global Division CIO for ABB’s Measurement and Analytics business. He brings more than 40 years of experience, including leading application and integration services at Information Control Corporation and supporting General Motors manufacturing operations through EDS. Episode 3: Building the Foundations for Scalable Industrial AI – with Jeff Witt of GE Vernova Guest: Jeff Witt, Global Process Digital Authority at GE Vernova Expertise: Manufacturing digital transformation, asset performance management, predictive maintenance, industrial AI. Brief Recognition: Jeff Witt is the Global Process Digital Authority for Manufacturing Assets and Maintenance at GE Vernova. He previously spent more than 12 years at Owens Corning in roles spanning global reliability systems, enterprise asset management, asset health, and digital transformation strategy. He holds a BS in Business Management from Cardinal Stritch University. Purpose-Built Industrial AI for Controlled Physical Operations Purpose-built industrial AI is designed for the equipment, safety requirements, and permitted actions of a specific production workflow. Antoine Bisson explains why this matters on the shop floor, where incorrect guidance can contribute to downtime, injuries, or fatalities. Those consequences make governance part of the system’s design from the outset. For Bisson, that means defining what information workers can access and how AI may answer their questions. Bisson argues that the test for releasing an AI feature is whether it delivers concrete value within appropriate safeguards. “It cannot be some fluff; it cannot be just AI, AI, AI. It needs to be tangible; it needs to be providing something that is concrete to the customer. Even if it takes more time to go to market, I’d rather do this and provide real value that is governed, that is safe, and safety is also a really key topic here.” — Antoine Bisson, CEO and Co-Founder at Poka Jeff Witt highlights that the design must also accommodate the equipment’s network and computing constraints. He describes shop-floor equipment deliberately isolated from external networks for security. For applications involving equipment or process control, he discusses running models closer to the equipment through edge computing rather than relying on cloud analysis. His example makes the architecture an application-specific decision about where data can travel and where a model can operate. Control also requires distinguishing between recommending an action and carrying it out. Bisson’s examples of creating an issue record, suggesting a possible cause, and changing equipment settings involve different permissions. He describes automatic equipment adjustment as a future scenario, conditional on a previously observed pattern and human confirmation that the proposed fix is appropriate. Together, these examples suggest three design questions for leaders evaluating industrial AI: Assistance: What information or recommendation will the system provide? Execution: Which workflow steps may it perform, and within what boundaries? Oversight: Which decisions require human validation, and how will that review occur? Connected Operational Data for Context-Specific Guidance Bisson treats context as a requirement for useful industrial AI. He includes information from machines and other shop-floor software, operators’ working practices, and recent safety incidents. After digitization, he argues, manufacturers need to establish whether their systems are connected and can communicate. Witt illustrates the operating context a frontline leader already brings to a decision during a machine alert: The machine and work cell. The people operating it. The asset’s history and recent events. Upcoming production plans. He contends that AI guidance must match this highly localized, machine-specific context to be trusted. A generic response to a worker’s first question can undermine confidence before the tool becomes part of daily work. Instead, recommendations should reflect the particular asset, product, and operating situation. Scot Burdette adds customer requirements to that context. In the manufacture-to-order work, Burdette describes specifications varying by order, and different people owning successive production stages. Each team needs to understand its role, access the relevant information, and pass that information along. He also connects access to information with training. Local teams need to prepare workers for their roles and understand whether they can find, interpret, and apply the data they need. His emphasis on cross-training, workflow modeling, and shared knowledge extends the requirement beyond connecting machine records: workers need to understand what the information means for their task. Burdette describes bringing operational and business data together so teams can use equipment information alongside production requirements when making decisions. The practical implication is that equipment information, production requirements, and working knowledge need to be usable together. An equipment manual alone may not explain the current order or the work completed upstream. Drawing on these points, leaders can trace one recurring problem from detection to resolution: Which records explain the asset’s history and current condition? Which work instructions and customer requirements apply? What training or expertise does the worker need to interpret the guidance? What information must reach the next team? Where does someone leave the workflow to find an answer? This exercise can reveal both missing system connections and gaps in the knowledge workers need to act on the information. Workflow-Embedded AI Assistance for Faster Frontline Problem-Solving Burdette emphasizes usability before performance measures because workers are unlikely to adopt a tool they find difficult to navigate or use. He also recognizes the difficulty of learning a new system while keeping pace with production and recommends time away from the line to become comfortable with it. He describes how workers can access device information, diagnostics, and troubleshooting guidance while working in the field. When workers need further support, they can share information and discuss the problem with a remote expert through the same system. This gives them both guidance they can consult themselves and access to someone who can help interpret an unfamiliar situation. Burdette connects this support to learning. Making experienced workers’ knowledge accessible can help newer employees understand problems and perform their roles more effectively. Local training and self-development opportunities remain part of that process. Another source of friction through issue reporting is addressed by Bisson. Instead of selecting an issue type, attaching an image, and entering details across several fields, he describes operators providing a voice description or picture that AI uses to interpret the problem. The intended change is fewer manual steps between noticing an issue and recording it for action. He also describes agents working beyond the initial report. In one proposed workflow, downtime exceeding 15 minutes triggers an issue record, investigation of possible causes, and information for a supervisor. This illustrates how agents could prepare a response while leaving consequential decisions subject to human validation. Witt approaches usability from the information workers must otherwise supply. He argues that an operator should not have to stop, open a chat window, and explain which machine they are standing beside. Assistance should already have the relevant work-cell and equipment context so it can offer guidance suited to the task. These examples shift the worker’s effort from navigating the interface to assessing guidance and deciding what to do next. Witt explains how human review continues within an AI-supported maintenance workflow. People still need to: Evaluate the recommendation. Plan and schedule the work. Arrange the necessary resources. Check tools, parts, documentation, and safety. AI supplies a starting point for assessment. The maintenance team retains responsibility for deciding how and when to act. The results of that work can also inform subsequent recommendations. Witt describes a thumbs-up or thumbs-down interface through which workers can indicate whether [truncated for AI cost control]