The companies that actually make things are quietly pumping the brakes on AI while software firms sprint ahead.
The Summary
- Honeywell's CTO says industrial customers demand 99.9999% accuracy while current AI models hit 85% — a gap that makes human workers cheaper than automation in many physical operations
- The reliability threshold for industrial AI isn't a technical challenge, it's an economic one: below six nines of accuracy, the cost of failures exceeds the cost of wages
- Manufacturing and industrial leaders are pushing back on the "AI will replace everything" narrative with actual deployment data from the field
The Signal
Honeywell CTO Suresh Venkatarayalu delivered what might be the most honest assessment of AI's industrial limits yet. His customers don't want 95% accuracy or even 99% accuracy. They want 99.9999%, and frontier models are landing around 85%. That's not a rounding error. That's the difference between a technology you can deploy and one you can't.
The math is brutal. In digital environments, an 85% accurate AI agent might miss some customer emails or generate a few bad marketing headlines. Annoying, fixable, rarely catastrophic. In a chemical plant or food processing facility, that same error rate means contaminated products, safety incidents, regulatory violations, and lawsuits. The cost of being wrong once erases the savings from being right a thousand times.
"Sometimes it's more expensive than having humans."
This isn't about AI being bad. It's about AI being exactly as good as it is, which isn't good enough for the physical world yet. The gap between 85% and 99.9999% represents thousands of edge cases, sensor failures, environmental variations, and the kind of embodied knowledge that experienced operators carry in their bones. A human worker knows when something sounds wrong or smells off. They know that this valve always sticks a little, that this reading spikes every morning when the sun hits the sensor, that this alarm can be ignored but that one can't.
Industrial companies are learning what software companies haven't had to face: reliability isn't a feature you can patch later. When Ecolab and Honeywell talk about AI being more expensive than humans, they're not being Luddites. They're being CFOs. The total cost of ownership includes the failures, and right now, the failure rate makes the ROI negative.
Key barriers to industrial AI deployment:
- Accuracy gap: 85% model performance vs. 99.9999% operational requirement
- Cost of error: Single failures can exceed total annual labor savings
- Domain specificity: General models lack the contextual knowledge humans gain from years on-site
The companies building AI for software workflows can afford to iterate fast and break things. The companies building AI for the physical world cannot. This creates a strange bifurcation in the agent economy. Digital agents will proliferate rapidly because the cost of their mistakes is low and the feedback loops are fast. Physical world agents will advance slowly, gated by reliability thresholds that are measured in nines, not percentages.
The Implication
If you're building AI agents, know which world you're building for. Digital workflows will see explosive agent adoption in the next 18 months. Industrial applications will move at the speed of trust, which is glacial. The companies that win in manufacturing and physical operations won't be the ones with the most advanced models. They'll be the ones that figure out hybrid systems where AI handles the predictable 85% and humans own the critical 15%.
For workers in physical industries, this is the rare good news in the automation story. Your job isn't safe forever, but it's safe until the models can hit six nines. That might be two years or it might be ten. Use the time.