Representative End User Clients
Representative Automation Clients
Representative Software Clients
In the first three episodes of this series, we rediscovered our industry’s foundational challenges, tracked the rise of what ARC calls the Industrial Data Fabric, and highlighted the domain-specific innovators driving much of the momentum. This journey has established the critical need for clean, organized, and AI-ready data as the bedrock for future innovation.
Now, we take the next logical step. In this fourth episode, we explore how all that AI-ready data gets activated. How do we move from data foundations to the new wave of reliable, agentic AI solutions? As I teased in my recent blog, “Industrial AI Needs Context Engineers, NOT Prompt Engineers,” this transition requires a deliberate and structured approach.
I was thrilled to welcome Vatsal Shah, co-founder and CEO of Litmus Automation, to discuss this critical link. Vatsal’s journey—from a hands-on industrial automation engineer frustrated by the challenges of integrating systems from Rockwell, Siemens, and Yokogawa to founding a company at the heart of Industrial DataOps—provides a unique and pragmatic perspective on what it takes to succeed.
Listen or Watch
For those who prefer to listen or watch, the full conversation is available here:
Listen on BuzzSprout:
Watch on YouTube:
For a deeper dive into the key takeaways from our discussion, read on.
Key Insights and Recommendations
1. Industrial DataOps Is Born from Real-World Plant Floor Frustration
The drive to create modern industrial data platforms stems directly from the deeply entrenched pain of integrating disparate, legacy OT systems. Before we can even think about AI, we must first solve the fundamental problem of unified data access and bring a modern, scalable approach to a world of proprietary protocols and decades-old control systems.
Vatsal Shah: “We had six engineers reverse-engineering code for four months just to get the data out. The very first line I wrote on a napkin before starting Litmus was, ‘We want to be Google for industrial data.’ The goal was to make it searchable, to make it awesome.”
Colin Masson: “You’re stirring up old memories from my time as a process control engineer. It really highlights the core issue: this is the legacy problem we have to solve first before we can successfully deploy and monetize new solutions on AI.”
2. “Context Engineering” Provides a Framework for Reliable Industrial AI
Moving beyond simple prompts to build reliable industrial AI requires a new discipline I call Context Engineering. In our conversation, Vatsal outlined a practical framework that demonstrates this is far more than just retrieval-augmented generation (RAG). It’s a structured methodology for providing AI with the dynamic information it needs to produce accurate, relevant, and deterministic outputs. This framework combines four key elements: the initial prompt, short-term memory of live events (e.g., current work order), tools to call external systems (e.g., SAP), and user preferences or guidelines that constrain the output.
Vatsal Shah: “We created a framework for context engineering with four key components: live ‘memory’ of what’s happening now, ‘tools’ to call external systems, access to live data, and the ‘user preferences’ or guidelines for the task. When you combine these four things, your AI output doesn’t just improve—it radically changes.”
Colin Masson: “In our world, it’s far more complex than a simple prompt. Understanding what’s happening on a factory floor is incredibly time-dependent and dynamic. We simply don’t have the luxury of large language models that have been pre-trained on that specific, real-time operational data.”
3. AI Agents Require a Hyperconverged Edge-to-Cloud Architecture
The long-term vision is a “swarm” of purpose-built AI agents orchestrating tasks across a manufacturing facility. However, the current reality is that factories are constrained environments—often air-gapped and lacking the abundant compute resources of the cloud. The pragmatic path forward is a hyperconverged infrastructure where applications and AI models are fluid, capable of running at the edge or in the cloud as the use case demands. This requires a broader toolbox than just generative AI; it includes machine learning, computer vision, and heuristics-based agents.
Vatsal Shah: “The vision is a swarm of purpose-built agents, each created to solve one problem consistently. For this to work, agentic applications must be completely fluid, able to run in the cloud one moment and at the edge the next, wherever they are needed.”
Colin Masson: “An ‘agent’ doesn’t necessarily mean ‘GenAI.’ Our industrial toolbox is much broader; agents can be based on machine learning, heuristics, or simple data retrieval. The key is that they can work as a collaborative process, running across multiple layers of the traditional automation stack.”
4. The Foundational Imperative: You Can’t Enable AI Without First Fixing Your Data
The immense potential of AI is finally creating the business case to solve our industry’s long-standing data challenges. Companies that previously invested in building a solid data foundation—collecting, cleaning, and contextualizing their OT data—are now positioned to rapidly deploy AI solutions. For them, enabling AI is not a multi-year project but is as simple as “turning on the switch” because the prerequisite work is already done.
Vatsal Shah: “Our customers who already invested in their data foundation sent us letters of appreciation, saying they would be first on the AI train simply because their data was ready. The question for every manufacturer is: are you going to keep patching together different vendors, or are you going to get the foundation done right now?”
Colin Masson: “We have to crack the data context code. That’s the only way for the entire enterprise to tap into the massive volume of data we generate in the factory—more data than the rest of the business combined.”
Diving Deeper
As we venture further into topics like building robust data infrastructures and modernizing architectures to effectively infuse AI, I highly recommend readers explore some of my existing research:
My blog series on "Assembling Industrial-grade Data Fabrics."
My series on "The Rise of A2A: Completing the Industrial AI Protocol Stack with OPC UA and MCP."
The article, "Core Capabilities of the Industrial-grade Data Fabric: Powering AI Infusion and Modernization," which is the sixth post in the Data Fabric series and delves into the essential solution services these fabrics enable.
These pieces offer a solid foundation for the themes Rick, Vatsal, and I have been unpacking in this series on Industrial Systems Engineering in the New Era of AI.
Stay Connected
The dialogue on industrial AI and digital transformation is constantly evolving. To stay informed and hear more insights from industry leaders, we invite you to subscribe to the ARC Advisory Group's Digital Transformation podcast series.
We believe the best conversations include diverse perspectives. If you are an innovator in this space and would like to contribute to a future discussion, please reach out to Colin Masson at ARC Advisory Group.
Engage with ARC Advisory Group
For ARC Advisory Group recommendations for Navigating the AI Wars, Closing the Digital Divide by Embracing Industrial AI, assembling your Industrial-Grade Data Fabric, and governing and guiding major decisions about enterprise, cloud, industrial edge, and AI software, please contact Colin Masson at [email protected] or set up a meeting with me, or my fellow Analysts at ARC Advisory Group.