COLUMN – In this column, three lean and technology experts respond to the same pressing question shaping today’s tech/AI debate. This month, we dive into why building people-first organizations in more important than ever.

Words: Theodor Panayotov, Marie-Pia Ignace, and Erasto Meneses

THE QUESTION

Jidoka was Sakichi Toyoda's way of ensuring technology took on what machines do well, so people could focus on what only people can do. More than a century later, organizations are again facing a choice: should technology replace humans or elevate them? As AI reshapes work at scale, how should leaders actively design organizations where people thrive rather than simply survive technological change? And what does Lean tell us about what that design looks like in practice? (By the way, this is the theme of this year's Lean Global Connection.)

THE ANSWERS

The story of Sakichi Toyoda's loom is usually told as a case of technology serving people. It deserves a more honest telling. The Type-G loom stopped itself when a thread broke — which meant one weaver could supervise dozens of machines instead of watching one. That invention made two futures possible: fewer weavers doing the same work, or the same weavers doing better work. The loom had no opinion on which. What became the Toyota Production System is what made the second choice systematic: freed human attention was deliberately channelled into problem solving, improvement, and the judgment no machine could perform. Elevation was never a property of the technology; it was a property of the design built around it.

We are at the same fork with AI, at far greater scale, and most organizations are choosing by default. They deploy AI into unchanged structures, and their people end up with the worst hybrid: supervising machine output they have no authority over, at machine pace, with human accountability.

Lean tells us what deliberate design looks like, and it is uncomfortably concrete.

Do what Sakichi actually did: separate machine work from human work, explicitly, role by role. Pattern detection, retrieval, drafting, and monitoring for the machine. Judgment, exceptions, relationships, and improvement for humans. Most AI deployments skip this step entirely; roles are left to erode rather than being redesigned.

Keep jidoka's direction of authority. The loom stopped and called a person; the abnormality flowed toward human judgment. The same must hold for AI: the person closest to the work must be able (and should be expected) to stop, challenge, and correct the algorithm. An organization where nobody can pull the andon cord on the AI has not automated intelligence. It has automated the conveyor and kept the surveillance.

Reinvest the freed capacity where people can see it. If the hours AI saves quietly convert into headcount targets, people will rationally resist and hide what they know. If those hours visibly become kaizen time, training, and customer contact, people will bring to the machine their best problems.

Finally (and this is what I see daily, building AI for nuclear, pharmaceutical, and defense organizations), treat expert judgment as the asset, not the cost. The tacit knowledge of a senior operator is precisely what AI cannot replace, and precisely what careless deployment destroys: through deskilling now, and through retirement without capture later. The design brief is to elicit, preserve, and scale that judgment — the problem we work on at Ethermind — and it is harder than any model.

Should technology replace people or elevate them? That is the wrong question. Technology will do whatever the organization around it is designed to do. Sakichi understood this a century ago. The leaders whose people thrive in the age of AI will be the ones who design for elevation on purpose, which is exactly the conversation on which this year’s Lean Global Connection will focus.

Should tech replace people or augment them? Put that way, this question invites a moral answer, and — at least in theory —everybody knows the correct one. It is also operationally useless, because no executive will openly announce an intention to replace people with AI, and no vendor will claim that their product will actually make the process worse.

In practice, there are four types of human–AI relationships. In the first, people merely connect broken systems; this is waste, not complementarity. In the second, they supervise the machine while it learns; this is useful, but temporary. In the third, AI handles standard cases, while humans absorb genuine variability and complexity. In this column, I'd like to explore the fourth type: when the machine must interrupt its own work and hand off the case the moment a defect appears.

This is where the lessons from Toyota become most relevant. Sakichi Toyoda's jidoka was never a moral choice about preserving human dignity. It was a device: a loom that stopped when a thread broke. Not a loom that wove better, nor one that simply alerted the operator, but one that refused to continue producing defective cloth. Toyota later called this autonomation: automation with a human touch. What would the equivalent be for AI?

There are two situations in which an AI system can fail. The first is at the start, by misinterpreting the situation it's dealing with — a problem of improving judgment, but not the one considered here. The second, more troubling, happens during execution: AI can make an elementary error without knowing or reporting it — misread an amount, match a payment to the wrong customer, or invent a refund policy.

In Moffatt v. Air Canada, a chatbot supplied a bereavement-fare rule that did not exist. Air Canada rejected the refund and argued that the chatbot was a separate legal entity responsible for its own statements. The judge did not accept the argument and held the company responsible for the information published through it. (The CanLII Blog)

How does this happen? A language model is probabilistic, not deterministic — repeat the same case twice and the results may diverge. Each AI action therefore carries an error probability. Consider a workflow of twenty successive actions performed at 99% reliability: end-to-end reliability drops to just 82%. At 99.5%, it's 90.5%. In most operational processes, that's nowhere near sufficient.

One naive approach would be to ask the model, or a second model, whether it thinks it's correct. But Sakichi didn't ask the loom whether it believed it was weaving properly. Since we can't fight the model's non-determinism, we need to find its digital equivalent: a thread that breaks, a weight that falls, a mechanism that stops.

Quote-to-cash is a long and critical process for companies, and it already contains such digital threads. An invoice must balance. A payment match must point to an existing receivable. A credit note must have a referenced cause. A payment schedule must stay within authorised terms. A customer with a zero balance must not receive a collection notice. These are not confidence scores — they are hard conditions: true or false, testable without consulting the model. Each can break, and each can stop the line.

This is the rule that the AI wave systematically violates: a model that verifies itself doesn't verify anything, and a second model that verifies the first shares its blind spots. The best systems will therefore need a dual architecture: AI to circulate and transform data, and a deterministic layer — the old information-systems way, ironically — to identify the defect and stop the work. Not a more reliable model, but a model whose errors are detectable by something other than itself.

Quality first.

AI is not asking organizations to choose between people and technology. It is asking leaders to redesign work so that both can perform at their best.

From a Lean Tech perspective, this is the real challenge. The question is not whether AI will replace people, but whether leaders are willing to redesign processes, decisions, and management systems around uniquely human capabilities.

Lean has always taught us that technology should amplify value, never become the value itself. Jidoka was never about removing people from the system; it was about freeing them from repetitive, predictable tasks so they could apply judgment, creativity, collaboration, and continuous improvement where it matters most. AI simply expands this principle to knowledge work.

Unfortunately, many organizations are still approaching AI as a technology deployment instead of an organizational transformation. They automate fragmented processes, accelerate existing waste, and celebrate productivity gains without questioning whether they are creating more value for customers or more meaningful work for employees. Faster waste is still waste.

Lean Tech offers a different path. It begins by understanding the problem before selecting the technology. It redesigns value streams before automating them. It develops people alongside digital capabilities. Most importantly, it creates management systems where humans and AI continuously learn from each other rather than compete with each other.

The organizations that will lead the next decade will not be those with the most advanced AI. They will be those capable of intentionally designing systems where AI enhances human potential, strengthens decision making, accelerates learning, and allows people to spend more time solving complex problems and creating value that only humans can deliver.

The greatest responsibility of a leader today is not to implement AI. It is designing organizations where technology continuously elevates people, and where people continuously improve technology. That is not only consistent with Lean Thinking, but it is the natural evolution of its purpose. In the age of AI, the most competitive organizations won’t be the most automated; they will be the most human-centered ones.