Never Let the Technology Precede the Problem

🏭 Breaking the Bottleneck is a weekly newsletter and interview series covering manufacturing technology and physical AI. Want to chat? Reach out at aditya@machinafactory.org or connect with me on LinkedIn.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.

"If you ask a veteran worker to write down everything they've learned over 40 years, the project will fail miserably. As humans, we have decades of experience, which makes us excellent at catching mistakes but not necessarily at documenting them."

You’ve held leadership roles at IBM, AWS, and now Google Cloud, each during a different phase of industrial digitization. What’s the single biggest lesson each stop taught you that the previous one couldn’t?

It’s fascinating to reflect on how rapidly industrial digitization has evolved throughout my career. My experiences at these three companies each taught me a unique lesson about transformation, with one building directly on the next.

My time at IBM was a masterclass in the sheer scale and complexity of enterprise transformation. When I joined to lead technology strategy, the market was flooded with point solutions, early cloud, mobile, and social vendors, each pitching their specific technology as a hammer looking for a nail. At IBM, we focused instead on the concept of convergence benefits. This was about looking at the holistic problem and working backward from the desired business outcome. The big lesson there was that enterprise transformation requires understanding how people, processes, and technology must intricately weave together to achieve that final result.

If IBM taught me how enterprises think, AWS taught me how they can execute with agility. AWS radically challenged the traditional approach of bringing in big teams to tackle a project all at once. Instead, they pioneered the “two-pizza team” philosophy—the idea that an effective project team should be small enough to be fed by just two pizzas. By keeping teams tight, you minimize lines of communication and eliminate the bureaucratic overhead that slows down decision-making. This emphasis on speed, nimbleness, and constant experimentation for the customer is exactly what fueled AWS’s explosive growth. By focusing on highly targeted, small-scale pilots that could be scaled rapidly, we proved that you don’t have to sacrifice velocity to solve enterprise-grade problems.

Now at Google Cloud, the lesson has been all about intelligence and the true power of data. As a data and AI company, our core mission is to organize the world’s data and make it universally useful, a mission that translates incredibly well to the manufacturing industry, which sits on a goldmine of both IT and OT data. In that data resides the narrative of everything happening within a company, why it’s happening, and what you can do about it. Google has made me realize that the ultimate phase of digitization isn’t just about understanding the enterprise blueprint or moving fast. It’s about listening to what the data is telling you, uncovering actionable insights, and making the entire enterprise fundamentally smarter.

One thread that runs through all of those roles is the IT/OT divide, something you’ve called the “enemy of profits.” After years of the industry talking about convergence, why is that wall still standing, and what does Google Cloud’s Manufacturing Data Engine do differently to tear it down?

The IT/OT wall still persists due to the inherent complexity of manufacturing. Unlike retail, for example, which only has to manage buying and selling, manufacturing operates on three distinct imperatives: first, you have to invent a great product, then you have to make it faster, cheaper, and better than the competition, and finally, you have to navigate a complex supply chain and logistics to sell it to the right people. This involves a massive, fragmented web of people, processes, and machines. Historically, the industry managed this complexity by using traditional standards like ISA-95, which explicitly separated shop-floor OT from corporate IT systems. Because these systems speak entirely different languages, the divide became deeply structural.

While that divide is starting to blur, Industry 5.0 and AI agents are completely changing the game. To an AI agent, data is just data; it doesn’t care whether information originates on the shop floor or in the cloud. The real challenge is the heavy lifting required to connect those machines, extract their data, and harmonize it. True harmonization means building a digital thread that can instantly tell you which customer ordered a product, which plant made it, during which shift, and on which specific machine.

This is exactly where Google Cloud’s Manufacturing Data Engine and Manufacturing Connect come in. We can seamlessly connect to more than 250 OT data sources, such as PLCs, HMIs, and SCADA systems, and automatically integrate that machine data with your IT systems of record, such as CRM or PLM platforms. By bridging these two worlds into a unified data foundation, we break down historical walls and give companies the single thread they need to build a truly intelligent enterprise.

A key piece of that convergence story is the Cortex Framework, designed to bridge heavy ERP data and SAP environments with real-time analytics. For a mid-market manufacturer running on legacy systems, what’s the honest timeline and investment to get from where they are today to a functioning “digital thread”? How should leaders benchmark whether the framework is driving measurable ROI?

To understand the honest timeline and investment required, you first have to look at how long this process traditionally takes. In manufacturing, you are constantly dealing with a “museum of machines”–hundreds of protocols and assets that are 30 to 50 years old and were never designed to talk to each other. Traditionally, a company would spend six to eight months just on data extraction, stretching the total project timeline out to 14 months. It’s a slow, frustrating process because if you ask five different legacy systems for today’s production numbers, you will get five different answers. They all have their own metadata and versions of truth.

Google Cloud Cortex Framework and our Industry 5.0 architecture fundamentally change this dynamic by acting as a translation layer. Cortex provides pre-built data models for complex environments like SAP, allowing you to connect legacy systems and extract data in days or weeks, rather than months. We sit on top of everything using a Unified Name space (UNS) with a pub/sub model, meaning all your legacy systems subscribe to a single source of truth rather than creating their own. By getting this data layer right, we compress the 14-month timeline into weeks. In fact, we’ve seen customers deploy more than 800 AI agents into production in under 10 months because their data foundation was finally seamless.

When it comes to benchmarking ROI, leaders shouldn’t view this as a multi-year gamble. Our global survey of over 500 manufacturing executives shows that generative AI is already delivering sizable returns, with 75% of leaders reporting immediate improvements in productivity. The most successful leaders are achieving these gains by following a “buy to integrate” philosophy. They don’t try to build everything from scratch, nor do they buy rigid, closed systems. They let Google Cloud handle the heavy data lifting, freeing them up to focus their time and budget on the specific use cases that drive real productivity and growth.

Even with the right data infrastructure in place, none of it matters without people to act on it. With 30% of manufacturing jobs going unfilled and 1.9 million projected vacancies by 2033, that’s a real problem. You’ve framed AI as a tool to capture retiring workers’ expertise and deliver it to new hires. How do you actually extract decades of tacit knowledge from a seasoned machinist and encode it in a way an AI agent can deliver reliably?

The manufacturing talent gap is undeniable, but the nature of the workforce is also undergoing a massive shift. The traditional image of a factory worker carrying wrenches is giving way to a new reality of “technocrats” who navigate the shop floor with tablets, configuring machines remotely and using computer vision to diagnose issues. Because of this shift, we can’t bridge the talent gap using old-school methods. If you ask a veteran worker to write down everything they’ve learned over 40 years, the project will fail miserably. As humans, we have decades of experience, which makes us excellent at catching mistakes but not necessarily at documenting them.

AI can play a profound role in capturing that hidden expertise by letting veteran workers validate information rather than author it from scratch. Instead of writing a massive manual, an expert can simply be recorded performing a complex repair. AI can instantly translate that footage into a step-by-step instruction guide or a synthetic training video. The veteran worker simply reviews and approves it, acting as the expert validator behind the AI.

With that verified data foundation in place, companies can deploy AR and VR tools to guide new hires through complex tasks in real time. The AI walks the technician through the process step by step, using automated validation to confirm each task is completed correctly before allowing them to proceed. Ultimately, this isn’t just about preserving legacy knowledge; it’s about dramatically compressing onboarding timelines and empowering companies to help generalist workers confidently perform highly specialized tasks.

You’ve spent two decades watching industrial tech waves: IoT, blockchain for supply chain, digital twins, and now agentic AI. Which of those earlier bets do you think the industry over-invested in, and what lesson from those experiences should manufacturers bring to agents?

Rather than a case of simple financial over-investment, it’s more useful to look at where the industry chased hype that ultimately under-delivered. From that list, blockchain and digital twins are the two clear examples where I believe the technology was applied to the wrong problems.

With blockchain, the technology was heavily hyped because people saw its success in crypto and started treating it like a hammer in search of a nail. They tried to force it into standard transactional supply chain problems without realizing it was built to solve specific multi-party trust issues, not to handle real-time, compute-heavy enterprise tracking. Similarly, early digital twins lacked a clear, unified definition. Many companies poured resources into building complex 3D CAD models of their shop floors just to have a visual representation on a screen, without tying those models to a specific operational problem or a measurable ROI.

The lesson for the era of agentic AI is clear: never let the technology precede the problem. Those earlier tech waves stalled because companies invested in the tool first and the use case second. With AI agents, the focus has to remain entirely on the business outcome. Leaders shouldn’t deploy an agent just because the capability exists; they need to identify a specific bottleneck and build the data foundation to solve that precise issue. When you work backward from the problem rather than chasing the tool’s hype, you ensure your technology delivers immediate value while protecting your transformation from reputational risk.

To contact Praveen, reach out to him on LinkedIn here. He’s always open to chatting and sharing valuable insights.

This newsletter is brought to you with support from our featured partners: AMT, IMTS, Jiga, Upkeep, and Industry 4.0 Club.