Digital Jidoka: Why Scaling Operations Demands Human Judgment and Intelligent Co-Pilots

A conversation with Marc Onetto — former Senior Vice President of Worldwide Operations and Customer Service at Amazon and GE Executive — on standardizing complexity, the "Mechanical Sensei," and how AI elevates frontline training and customer empathy.

By Louise Servoin · 2026-06-15 · 14 min read

## Marc Onetto – Principal, Leadership From The Mind And The Heart LLC | Former EVP of Worldwide Operations, Solectron | Former VP and GM of Global Supply Chain, GE Medical Systems.

Expert in Lean-Six Sigma integration, global logistics, and high-tech supply chain operations. École Centrale de Lyon and Carnegie Mellon University (Tepper School of Business) alumnus. Based in Seattle, Washington, United States.

Few executives have shaped the backbone of modern global commerce quite like Marc Onetto. From driving Lean transformations across General Electric’s complex medical manufacturing in the 1990s to restructuring Solectron’s massive high-tech manufacturing footprint, Onetto’s true trial of fire came in 2006. Recruited by Jeff Bezos’s team, he was tasked with scaling Amazon’s fledgling logistics network into a multi-billion-dollar global juggernaut.

Under his leadership, Amazon’s fulfillment footprint grew tenfold, powered by a revolutionary synthesis of Lean manufacturing principles and cutting-edge software engineering. We sat down with Marc to discuss the evolution of Jidoka (autonomation), the role of Amazon's pioneering "Mechanical Sensei" algorithm, and how the modern frontier of Artificial Intelligence is reshaping how we train, empower, and co-pilot frontline operators.

## The Philosophy of Autonomation

Marc, you have often described yourself as a "Lean lunatic". Yet, you spent your career at the vanguard of high-tech companies like GE, Solectron, and Amazon. How do you reconcile traditional Lean philosophy—which many view as purely manual and physical—with highly automated digital ecosystems?

It is a common misconception that Lean is anti-technology. The definition of muda (waste) is simple: it is any activity that the customer would refuse to pay for if you showed it to them. That definition is precisely why Jeff Bezos and I aligned immediately when I joined Amazon. He wanted to build the most customer-centric company in the world, and I knew that Lean was the vehicle to systematically eliminate everything that did not bring value to the customer.

The core of Lean isn't about keeping things manual; it is about Jidoka. Jidoka is the foundational principle of stopping the line the instant a defect is detected. Sakichi Toyoda's automated loom stopped the moment a single thread snapped—it didn't replace the human; it prevented defects and freed the human from mindlessly watching a machine.

Autonomation goes one step further: it leverages digitalization and robotics to automate the non-value-added tasks where humans don't bring any added value, while preserving that same safety stop the moment a defect occurs.

In a digital environment, this is even more critical. If you automate a bad process without human-centric Jidoka, you simply accelerate the generation of waste. Technology must serve as a co-pilot, not a replacement. Whether we are talking about physical factory floors, customer service software, or modern AI systems, the goal is the same: automate the repetitive, low-value, Three-Sigma tasks so that a highly flexible, creative human being can elevate the entire process to Six Sigma.

## Redesigning the Digital Shop Floor

At Amazon, you pioneered the "Digital Andon Cord". How did you translate a physical rope from a Toyota assembly line into a digital customer service workflow?

In traditional operations, customer service agents are treated as passive record-keepers. They hear complaints, click through rigid screens, but have zero authority to fix the root cause. That is incredibly demoralizing.

We changed that by introducing the Digital Andon Cord. The first time we used it was a perfect illustration. A customer called to say they had received a hard disk instead of the headphones they ordered. The agent entered the complaint, and our system cross-checked the defect database. Because we don't stop the line on a single incident—it could be a one-off mistake—we waited for the second report. When it came in, the system told the agent: this has happened before, stop the line. With one click, the agent suspended the sale of that product, rendering it unbuyable, and pulled the remaining warehouse inventory to our Quality Lab.

What we found was fascinating: a factory in China was producing both headphones and hard disks, and the barcodes had been swapped on the packaging. Because every operation in the fulfillment center is barcode-driven, the pickers had no way of catching it—they scanned, the system said correct, and off the wrong product went. We then went back and pulled every single mislabeled unit from every distribution center in North America.

Every two weeks, the retail team and operations reviewed every Andon pull together, because the goal is never to stop the line for its own sake—the goal is to restart it as quickly as possible, having understood and eliminated the root cause.

Our retail colleagues were initially terrified of losing sales, but Jeff Bezos fully backed us, because we trusted our people. Frontline workers are the best source for identifying problems—less than 2.5% of those Andon pulls were unnecessary, and we eliminated 50,000 to 100,000 defects annually.

But the most underrated impact of the Digital Andon Cord was on the agents themselves. Customer service is, by nature, a thankless job: customers never call to say I love Amazon; they call because something went wrong. Giving the agent the power to look the customer in the eye and say, I'm sorry this happened to you, but thanks to your call, I just pulled this product from the catalog—no other customer will face the same issue, completely transforms the dynamic. It gives the customer immediate proof that Amazon takes them seriously, and it gives the agent a pride and a job satisfaction that simply didn't exist before. That's the human side of Jidoka—and it is at least as important as the technology behind it.

"Automation is not about replacing human judgment; it is about automating the low-value steps so a highly flexible human being can operate at Six Sigma." — Marc Onetto

## The Transition to Machine Reasoning: The "Mechanical Sensei"

Long before the current Generative AI wave, you deployed sophisticated algorithms at Amazon, most notably an internal software known as the "Mechanical Sensei". What was the role of this system, and how does it relate to the AI tools we see today?

The "Mechanical Sensei" is a prime example of complex machine reasoning. In our fulfillment centers, logistics are highly complex. We don't organize inventory like a traditional warehouse—books aren't neatly stacked next to books, because it is too easy for a human picker to grab the wrong edition of the same title.

Instead, our inventory is placed semi-randomly. The "Mechanical Sensei" acts as the invisible hand orchestrating this symphony. It uses logic-based algorithms, digesting vast streams of metrics—inventory levels, delivery lead times, warehouse capacities, and promised delivery dates—to minimize shipping prices and simulate optimized workflows across our entire global network. It tells the worker exactly what route to take to pick an item.

This is the precursor to the AI-driven systems of today. But the lesson we learned from "Mechanical Sensei" is that algorithms must remain explainable and supportive of standard work. If the algorithm makes decisions in a black box without standardizing the physical worker's workflow, you fail.

For instance, when I worked on the stow line myself, I realized I couldn't meet my standard productivity target because my scanner battery was weak and I had to scan items four times. The algorithm didn't know that. We had to run a physical Kaizen event to identify the root cause and build a process to monitor scanner batteries. AI can optimize the macro-parameters, but the human must still validate and refine the micro-realities on the Gemba (the actual place where work happens).

## AI-Assisted Training: Elevating Human Empathy

One of the most exciting frontiers today is AI-assisted training and learning. How can AI act as an educational "co-pilot" for frontline operators, and how does that connect to your work in reducing process waste?

This is where we see the most profound synergy. When we mapped our customer service workflows, we discovered that an associate had to navigate through 29 different software screens to resolve a single, standard customer query. That represents massive muda—not just of time, but of cognitive energy. We ran Kaizen events to redesign the tools and slash that down to 8 clicks.

Now, imagine overlaying modern Generative AI and AI-assisted learning onto that environment. Instead of forcing a newly hired customer service agent to spend weeks memorizing complex operating manuals, corporate hierarchies, and 29 different legacy databases, an AI-assisted co-pilot can act as a personalized, real-time training assistant. It can instantly retrieve the aggregate knowledge of thousands of past Kaizen events and surface the precise, contextual standard work instructions an agent needs in the exact moment of customer friction.

This completely revolutionizes onboarding and continuous education. We shift from a rigid, one-size-fits-all training model to a truly personalized, "n-to-1" learning model. But the metric of success for AI-assisted training must not be "how fast can we get the human out of the loop." The metric must be "how much cognitive load did we remove from this associate so they can focus entirely on the customer?"

Our number-one customer service tenet at Amazon was simple: treat customers like friends or family. If your associate is stressed out trying to navigate rigid software interfaces or struggling to recall training protocols, they cannot deliver empathy. By utilizing AI-assisted training to act as an operational co-pilot, we accelerate the learning curve, build baseline competency in hours rather than weeks, and free up the human mind to do what machines can never do: connect with another human being.

## The Human Core of Innovation

We are seeing a significant push toward total automation. Is there a danger in relying too heavily on AI as a decision-maker rather than a decision-accelerator?

Absolutely. If you build a system that removes the need for humans to think, you have designed a failing system. Behavior follows structure. If your system optimizes strictly for convenience and algorithmic decision-making, you create a culture of intellectual passivity and dependency.

Let me give you a concrete example of what this looks like in practice. I currently sponsor research at Carnegie Mellon University on what I call intelligent operations—the merging of Lean thinking with modern technology. One of the questions we are working on is deceptively simple: what is the defect rate of a robot? We know that humans are essentially three-sigma machines, we make errors roughly 6.6% of the time. That's why Lean invented tools like Poka-Yoke, the error-proofing mechanism that physically prevents a mistake from happening. The electrical wall socket is the perfect example: you cannot plug it in the wrong way, regardless of how distracted you are. It turns a two-sigma operator into a six-sigma one for that specific action.

But what about robots? Robots make mistakes too. The most famous robotic error in 2026 is called a hallucination—a very elegant word for the machine that got it wrong. Self-driving cars make mistakes. Painting robots make mistakes. And here's the catch: when a robot fails, it fails fast, and at scale. If a painting robot mis-paints one door, it will mis-paint a hundred before anyone notices at the end of the line.

A real example: when I was on the board of Flex—the contract manufacturer that acquired Solectron—we tackled this exact issue on PC board assembly lines. These lines have about thirty stations placing components onto silicon boards, then soldering them. For years, we relied on human inspectors at the end of the line to catch defects—a soul-crushing, low-value job. We replaced that by installing AI-powered image recognition cameras at every single station. The moment a component is misplaced, a solder joint is wrong, or the wrong part is loaded, the system stops the line at that station—not at the end. In effect, it is a robot supervising another robot. That is Jidoka with artificial intelligence: the Andon Cord of the digital age.

But, and this is the critical point, the AI can detect the defect and stop the line. It cannot tell you why the defect happened. That requires a human being to walk the Gemba, ask “why” five times, and re-engineer the process so the defect never happens again. The robot will then learn from that root cause and never repeat it. That is the proper division of labor between AI and humans.

AI can find patterns in a billion data points, and tools like "Mechanical Sensei" can optimize our shipping routes. But AI cannot sit in a room, look an operator in the eye, and understand the cultural resistance to change. It cannot feel the weight of a strategic pivot. Judgment and operational empathy are uniquely human traits. The future of operations isn't about choosing between human hands or artificial intelligence; it is about building an architecture where AI trains, guides, and co-pilots the human, but the human always remains the heart and mind of the operation.

## The Golden Thread

We often see a gap between what technology can do and what professionals actually adopt. How do we ensure AI becomes a 'co-pilot' that elevates human judgment rather than a tool that encourages intellectual passivity?

The answer is in how you architect the division of labor. We can delegate the act of production to robots—and increasingly, we should. We can even delegate the detection of defects to AI, by using one algorithm to supervise another, the way we do with image recognition on a factory line. That's the new Digital Andon Cord: a machine watching a machine, ready to stop the line.

But the moment a defect occurs, you absolutely need a human being. Because if the algorithm could have identified the root cause, the defect would never have happened in the first place—the system would already have learned to prevent it. Root cause analysis, contextual judgment, and the capacity to redesign a process so the same mistake never repeats: these remain irreducibly human. The robot executes. The AI watches. But the human teaches the system not to fail again.

If you build a culture where AI co-pilots the human—where it absorbs the cognitive load of low-value tasks and frees the human to think, investigate, and improve—you create the conditions for true continuous improvement. If instead you build a culture where AI replaces the human's need to think, you create intellectual passivity. Behavior follows structure. Design your systems to make humans smarter, not redundant.

Tags: Operations, Lean Manufacturing, AI, Customer Experience, Logistics, Leadership, Amazon, Continuous Improvement