Title: The PoC Graveyard: The 100x Rule of Factory AI
Don't Automate Chaos. I spent years at Sysmex and JTEKT—here is why most AI pilots fail before they start.
I have walked through a graveyard. It is the graveyard of “Proof of Concepts” (PoCs) in Japanese manufacturing.
Having worked at Sysmex (medical devices) and JTEKT (Toyota Group), I have seen the same tragedy repeat itself. A young new business planner—smart, enthusiastic, but inexperienced—comes up with a “cool AI idea.” They write a beautiful plan on a slide. They launch a PoC. And then, it dies.
Why? Because they skipped the most important rule of the “Gemba” (the actual site).
Here is the truth about why AI fails in factories, and the “100x Rule” you need to know.
The Trap: The “Desktop” Planner
The biggest cause of failure is simple: The planner doesn’t know the process. They plan solutions based on “what might work” without ever stepping onto the factory floor. They try to apply AI to a process they haven’t visualized.
In manufacturing, we make the same product thousands of times. Consistency is everything. If you don’t know the current state (”As-Is”), you cannot define the ideal state (”To-Be”). Without understanding the “Why” of the current workflow, any “Solution” is just a guess.
The Golden Rule: Standardization Before Automation
In the Toyota Production System, we talk about “Standardization” (Hyojunka). We train workers to be “Multi-skilled” (Tanoko), ensuring that anyone can do the job the same way.
But be careful: Standardization isn’t paperwork. It’s reducing variation so the same input produces the same output—regardless of who’s on the line.
If you ignore this and try to inject AI into a messy, human-dependent process, here is the formula I discovered:
Automating a Non-Standardized Process = 100x Chaos
Automating a Standardized Process = 100x Profit
Note: “100x” isn’t a scientific constant—it’s a field rule of thumb. When you speed up variance, you scale failure faster than you scale output. AI is an accelerator. If you accelerate a bad process, you don’t get efficiency. You get a disaster at high speed.
The Checklist: Factory AI Readiness
If you are trying to introduce AI to a factory, stop coding. Instead, go to the Gemba and ask these 5 questions. If you can’t check all of them, you aren’t ready for AI.
Factory AI Readiness Checklist
✅ Can you draw the workflow in 5 boxes? (If it’s too complex to draw, it’s too complex to automate.)
✅ Is there a stable definition of “good” vs “bad”? (AI needs ground truth.)
✅ Does the process produce consistent inputs? (Garbage in, garbage out.)
✅ Can three different workers produce the same result? (This tests standardization.)
✅ Are the top 3 failure modes known and logged? (You can’t fix what you don’t track.)
Conclusion
We often think AI will solve our problems. It won’t. AI only amplifies what is already there. If your Gemba is efficient, AI makes it legendary. If your Gemba is chaos, AI makes it a catastrophe.
Before you build a model, build a process.
"The planner doesn’t know the process." And when that happens the do all the stuff you mention like scaling broken systems as one example. Musk put this very well in those 5 rules he outlined a couple of years ago.
Musk’s 5-Step Engineering Algorithm:
1. Make Requirements Less Dumb: Question every requirement, especially those from smart people, as they can be the most dangerous if incorrect. Every requirement should have a specific person's name attached to it; not just a department.
2. Delete the Part or Process: If you are not adding parts back in at least 10% of the time, you are not deleting enough. The best part is no part.
3. Simplify or Optimize: This step is only to be done after the first two steps. A common error is optimizing a thing that should not exist.
4. Accelerate Cycle Time: Speed up the process, but only after streamlining to avoid doing the wrong thing faster.
5. Automate: Automate only after steps 1-4 are completed to avoid automating a flawed process.
So yes, if you amplify non-optimized workflows you end up overwhelmed very quickly with all the amplified mess and your dreamy 1.3x to 100x improvements go *poof*
As a note, I always push teams I'm supporting to to diagrams as code these days; usually using mermaid and sequence diagrams. Total game changer! But, for sure we would struggle to get the to five boxes. I suppose that might mean we need to do a better job at decomposing the system.
Great post, thanks for sharing over on my post. :)
AI doesn’t create efficiency. It amplifies whatever process you already have.
In factories, that means one rule: Standardize first — or scale chaos 100×.