The AI Rule Toyota Figured Out Decades Ago

Most of the business owners I talk with feel steady pressure to move faster on AI. Competitors are announcing new tools, vendors are calling every week, and somebody on the leadership team has usually read an article that makes it sound like everyone else is already done. So the conversation tends to start with how quickly something can go live.

I understand that pressure, and I'd never advise an owner to sit still while the technology moves. But I've come to believe the more useful starting question is how you'll know you're ready, and who's allowed to stop the rollout if you aren't. The best answer to that question I've ever come across comes from a car factory, and the idea behind it is older than most of the people working in AI today.

The cord anyone could pull

The Toyota Production System is one of the most studied ways of running an operation anywhere in the world, and one of its founding ideas is something Toyota calls jidoka. Loosely translated, it means building the ability to detect and stop for problems into the process itself, so that defects get caught where they start.

The best-known example of jidoka is a rope. Toyota's own 1992 description of its production system explained that a worker anywhere on the assembly line who notices an abnormality can stop the production flow by pulling a cord overhead or pressing a button. It's usually called the andon cord. When someone pulls it, a board lights up to show exactly which station has the problem, and a team leader comes over to help. If the two of them can't solve it within that station's work cycle, the line stops.

If you've spent your career keeping a line moving, that can sound a little crazy. Giving every worker on the floor the power to halt production seems like an invitation to constant interruptions. What made it work at Toyota was the culture surrounding the cord. In the classic telling of the system, the team leader who arrived at the station thanked the person who pulled it, and the problem became the priority. Workers learned that raising a problem early was valued, so they raised problems early. Small defects got fixed at the station where they began, before they could turn into a thousand flawed cars at the end of the line. Over time, the ability to stop made Toyota faster, because mistakes stopped compounding.

A recent essay on The Lean Blog made an observation about the cord that I think applies directly to AI. Installing the rope was the easy part. Whether anyone actually pulled it depended on everything around it: how supervisors reacted, whether the person who pulled it got thanked or blamed, and whether the problem truly got fixed. A cord that nobody feels safe pulling is decoration.

Most AI rollouts have no cord

Now think about how a typical AI rollout goes at a small or mid-sized company. A tool gets chosen, often because the vendor's demo was impressive. It gets switched on, and people are told to start using it. What's usually missing is any shared understanding of what "working" looks like, what "not working" looks like, and who has the authority to say the tool isn't ready.

So when the AI starts producing poor results, nobody pulls the cord, because there isn't one. The bookkeeper notices the transaction categories are off, assumes someone else is handling it, and quietly fixes them by hand every week. The dispatcher stops trusting the schedule suggestions and goes back to the whiteboard. The tool keeps running, the workarounds pile up, and six months later the leadership team concludes that AI didn't work for their business, when what actually failed was the process around it.

The research suggests this pattern is common. In 2024, RAND, the nonprofit research organization, interviewed 65 experienced data scientists and engineers to understand why AI projects fail. Its report notes that by some estimates, more than 80% of AI projects fail, roughly twice the failure rate of technology projects that don't involve AI. The most common root cause the interviewees described was strikingly ordinary: the people involved misunderstood or miscommunicated what problem the AI was supposed to solve. Other causes included not having the right data and chasing the latest technology instead of the actual business problem.

RAND's interviews focused mostly on larger, more technical projects than a typical small business would take on, so I wouldn't read that 80% figure as a forecast for your next rollout. The leading cause, though, applies at any size. When nobody can say clearly what a tool is for, nobody can say clearly when it's failing, and that's exactly the condition under which nobody pulls the cord.

The companies building AI are pulling their own cords

One of the more striking things about this moment is that the companies building AI have started stopping their own lines, in public. At the end of September, OpenAI called off the planned October release of a new model after it fell short of the company's own safety standards in testing, and the company has also paused the training that lets its most capable models use outside tools after a series of incidents with its test systems. You can debate the details and the motives, and plenty of people are. The principle is one Toyota would recognize immediately: when the result isn't right, you stop and fix it before it reaches the customer, even under real pressure to keep going.

If a company with billions of dollars riding on speed can pause to get it right, a 50-person business can certainly take a few weeks to make sure it's ready.

"We can't afford to slow down"

The strongest objection to all of this is cost. Every week spent preparing is a week without the benefit, and stopping a tool that's already live feels like admitting the money was wasted. I take that seriously, because for a small business, time and cash really are tight.

Toyota's experience points the other way, though. The andon cord looked expensive on any given afternoon, since every pull cost a few minutes of production. Across months and years, it made the whole system faster, because problems were fixed once, at the source, instead of being worked around forever. AI rollouts follow the same economics. A few weeks of preparation, plus a team that feels free to raise problems early, costs far less than a tool that runs badly for six months while people quietly route around it.

There's a less visible cost to skipping readiness as well. When a team watches an AI tool get switched on, produce poor results, and keep running anyway, they learn a lesson about every tool that comes after it. They learn to nod along and work around it. The next rollout starts with a skeptical team, and the one after that starts with a cynical one. A rollout that began with clear goals and a real cord builds the opposite habit, and people trust the next tool a little faster because they've learned to trust the process.

What ready looks like

"Readiness" can sound abstract, so it's worth describing concretely. Before an AI tool or agent goes live on real work, I'd want a team to be able to answer five questions, and none of them is technical.

The first is what specific job the tool is doing. That means naming the task, the people it affects, and how the work gets done today. "Speed up invoicing" is a goal. "Draft invoice line items from completed work orders for the office manager to review" is a job you can actually measure.

The second is what good looks like in numbers. Pick one or two measures you can check, such as hours spent per week, error rate, or turnaround time, and write down where they stand before you start. Without a starting point, nobody can say later whether the tool helped.

The third is what "not working" looks like. Decide in advance which signs would make you pause, such as wrong numbers going to customers, a growing pile of manual corrections, or people quietly avoiding the tool. Writing those signs down ahead of time means nobody has to argue about them later.

The fourth is who can pull the cord. Name the person with authority to pause the tool, and make it clear to everyone using it that raising a problem is welcome. Then, the way Toyota's team leaders did, thank people when they speak up.

The fifth is the fallback. Know how the work gets done if the AI is paused tomorrow. If nobody remembers the old process, pausing becomes so painful that the cord never gets pulled.

These questions are best answered by the people who do the work every day, which is one more reason we think of AI adoption as a people project first.

Where Avolis stands

Readiness is where every one of our engagements begins. Our process starts with diagnosis, because we want to understand where the work in your business actually gets stuck, what good looks like for your team, and what could go wrong, before anything gets switched on.

From there, we choose tools that fit and explain why, and we build the workflows and agents around a specific job with a measurable result. Before anything goes live, we agree with your team on what "not working" looks like and who can pause it, and we make sure the old process is written down so pausing is always an option. Then we train each role, capture what your best people know, and stay involved through the weeks after launch, which is when most rollouts either take hold or quietly fall apart.

Toyota became one of the most efficient manufacturers in the world in part by making stopping safe, quick, and normal. If you'd like an honest look at whether your business is ready for its next AI project, and what it would take to get there, a good place to start is our AI Readiness Evaluation.

Sources

• Lean Blog, "What Toyota's 1992 Booklet Says About Stopping the Line, and What It Leaves to You" (June 2026): https://www.leanblog.org/2026/06/toyota-line-stop-cord-1992/

• IT Revolution, "The Andon Cord": https://itrevolution.com/?p=68086

• RAND Corporation, "The Root Causes of Failure for Artificial Intelligence Projects and How They Can Succeed" (August 2024): https://www.rand.org/pubs/research_reports/RRA2680-1.html

• Asia Financial, "Study Suggests Ways to Overcome High Failure Rate in AI Projects" (August 2024): https://asiafinancial.com/study-suggests-ways-to-overcome-high-failure-rate-in-ai-projects

• The Next Web, "OpenAI cancels October launch of GPT-6.1 Astra after failed safety tests" (September 2026): https://thenextweb.com/news/openai-cancels-launch-of-gpt-6-1-astra