The AI Productivity Gap: Why 1890s Factory Owners Had the Same Problem as You (and the Secret to GenAI ROI)

Why electricity failed at first

In the late 19th century, factory owners were told that electricity would change everything. The new technology promised a new era of prosperity through lower costs.

Many bought in early, investing large sums to rip out their huge coal-fired steam engines and replace them with equally huge electric motors.

And then, for nearly thirty years, almost nothing happened.

The new machines were cheaper to operate and easier to maintain; however, productivity didn’t skyrocket.

And so, to the frustrated CEOs of the 1890s, electricity looked like just another expensive disappointment. This historical stagnation contains a key lesson for anyone leading AI adoption today.

But to understand what happened , we first need to understand how a steam-powered factory worked.

The productivity boom of distributed power

Productivity is a property of the system, not just the tool; swapping the engine is useless if you don’t cut the belts.

When electricity first arrived, managers made a one-to-one replacement, swapping the steam engine for an electric motor, keeping the line shafts and rigid floor plan.

This was, of course, cleaner and much more convenient than a coal-fuelled steam engine. But it didn't solve the underlying friction of mechanical power distribution. It wasn’t until the 1920s that managers stopped seeing electricity as a ‘better steam engine’.

It was then that they finally understood that electricity’s potential was best used in the form of distributed power. This paved the way for the unit drive: as electric motors became smaller and more capable, their applicability increased, which in turn drove further efforts to shrink them until they could be dedicated to individual machines.

Once the "belts" were cut and the ceiling was clear, factories were finally reorganized into assembly lines based on the flow of materials. Productivity suddenly boomed, because the technology allowed the environment to be flexible.

Don’t do AI adoption like a 1890s factory manager

True revolution requires the courage to clear the ceiling - dismantling the rigid hierarchies and synchronous dependencies built for a pre-AI world.

Today, most supply chain organizations are using AI exactly like a 1890s factory manager used a giant electric motor. They are dropping powerful large language models into legacy environments, slapping a chatbot, copilot, or agent on top of twenty-year-old procurement processes and rigid 9-to-5 cultures. This is, in essence, using a super-motor to turn rusty leather belts.

In a recent webinar hosted by Celonis, supply chain influencer Lora Cecere called this ‘AI Stupid’: overlaying AI onto existing supply chain workflows without rethinking them.

For example, current use of AI often requires a human to manually ‘crank’ the system with prompts at every step, keeping us firmly in the mechanical age. Even more elaborate AI deployments face the same ‘invisible belts’, as IDC’s Stephanie Krishnan explained: “[think of] building an agentic workflow that can rebalance production autonomously and the operations team insists on approving every output manually.”

What’s more, the line shaft system that’s been used so far in supply chains is coming to its limits. Supply chains are dealing with more volatility and higher complexity than ever before, hitting a cognitive ceiling of what humans can manage and comprehend.

We’re currently in the big motor phase of AI, using advanced technology to power old, inefficient workflows.

The key to the AI productivity boom: multi-agent systems

If the steam engine was the central model and electricity was the distributed grid, the unit drive of the AI era is the autonomous agentic workflow.

Just as the 1920s unit drive succeeded by being small, specialized, and decentralized, the AI boom will happen when we shift toward multi-agent systems. These are micro-agents specialized for specific tasks - research, drafting, or fact-checking - that talk to each other autonomously without needing a central human manager to prompt every step.

This means that instead of implementing agents haphazardly, without rethinking workflows, supply chains redesign their ways of working to make the most of these orchestrated multi-agent systems. Each micro-agent can work at its full capacity, and its potential for autonomy (within carefully designed guardrails) is fully leveraged.

We are also seeing a shift toward Small Language Models (SLMs) that can live locally on devices and Vertical AI trained exclusively on industry-specific standards. These specialized "motors" are 100x more efficient because they don’t waste energy on general tasks like writing poetry when they are built to weave the fabric of a specific industry.

The Unit Drive of AI consists of small, specialized, autonomous agents that "sew the fabric" of supply chain processes while we sleep.

AI can’t work effectively without the right context

However, agents cannot function in a vacuum; to remove the mechanical constraints of the digital office, you need a map of the floor. This is where context plays its critical role. If Agentic AI is the unit drive, the context of your supply chain is the standardized voltage, wattage, and plug-in components required to ensure compatibility, safety, and ease of maintenance. In order to be effective, AI agents need the operational context of a business: it understands the relationships between all the documents, materials, and people that make up your supply chain and how they’re interconnected and interdependent.

With this context, your agents can be orchestrated following the rules of your business, so that hundreds of agents don’t work at cross-purposes. Leaders stop operating the machine and start managing the line.

In an AI-native organization, the factory floor changes completely. Work becomes asynchronous by default, happening 24/7 as agents turn the machines while humans shift to the role of floor manager inspecting the final output for quality control.

The need to click through ten different apps disappears as AI flows data between points automatically. We’ll know we’ve reached the unit drive era when we stop talking about AI altogether. Just as people in 1930 stopped saying they were using an electric-powered sewing machine, and just said they were sewing, the productivity explosion will happen when the chatbot interface disappears, and AI becomes deeply embedded in the task itself.

How to redesign the factory

Don’t just buy a better engine; use the Celonis Platform to redesign the factory.

Want to start moving your supply chain towards the unit drive era of AI? The Celonis Platform brings together process data, business knowledge, and intelligence from all your systems, applications, and devices to create a dynamic, real-time digital twin of your operations.

It combines hindsight, insight, and foresight to give your people and your AI agents the operational clarity to reason correctly, decide sensibly, and act reliably.