Caterpillar's AI rollout: the workflow is the hard part
Caterpillar has automated mining for decades and $100M is going to retraining, not models. The AI deployment bottleneck is the jobsite, not the algorithm.
Caterpillar has been running autonomous haul trucks in mines for decades. It has 1.6 million connected assets and more than 16 petabytes of structured data. If anyone had earned the right to say the model is the hard part, it is them. Their CTO says the opposite. The AI deployment bottleneck is the workflow.
What actually happened
TechCrunch reported on August 30 on how Caterpillar is applying its mining-automation experience to a company-wide AI rollout. The projects are unglamorous: a Cat AI Assistant that answers repair questions by voice for field technicians, site scanning and digital twins, legacy code modernization, and internal software development.
CTO Jaime Mineart's framing: "The hard part about autonomy and about physical AI is incorporating that technology into the customer jobsite and into the workflows."
The spending backs it up. Caterpillar is putting $100 million over five years into workforce training across roughly 118,000 employees — moving operators from running one machine to supervising several remotely, and using experienced operators to train the systems rather than replacing them first and figuring it out later.
This is happening from a position of strength, not desperation. Q2 2026 sales and revenues hit $20.5 billion, up 24% year over year and the first quarter above $20 billion in company history, with Power Generation sales of $3.098 billion, up 29%, largely on data center demand.
Why this matters for your business
You have five to five hundred employees, not 118,000. The lesson scales down cleanly, because it is not about size.
Every failed automation project we have inherited had a working model in it. The model summarized the ticket, drafted the quote, read the invoice. It worked in the demo. It died because nobody changed what happens next — who reviews the output, what the exception path is, which screen the tech actually has open when their hands are dirty. Caterpillar's AI assistant is voice-driven for a reason. That is a workflow decision, not a model decision.
The other transferable piece is the training line. Caterpillar is spending on retraining before it is spending on replacement, and it is using its best operators as the teachers. If you are rolling out an AI tool and your plan for adoption is a Loom video and a Slack announcement, you have budgeted for the model and not for the thing that actually determines whether it gets used.
Pick one recurring workflow. Map who touches it and where the handoffs are. Then decide where the model goes. In that order.
Key takeaways
- Caterpillar runs 1.6M connected assets and 16+ PB of structured data, with decades of autonomous mining equipment in the field
- CTO Jaime Mineart: the hard part of physical AI is fitting it into the customer jobsite and the workflows
- $100M over five years is going to workforce training across ~118,000 employees, using experienced operators to train the systems
- The Cat AI Assistant is voice-first because field techs have their hands full — a workflow decision, not a model one
- Q2 2026 revenue was a record $20.5B, up 24%, with Power Generation up 29% on data center demand
Your automation pilot didn't fail because the model was bad. We map the workflow and the handoffs first, then put the model where it actually removes work. See what we've shipped, or tell us which process is eating your week.
Sources: TechCrunch, CNBC.
- #ai-adoption
- #physical-ai
- #automation
- #workflow
- #change-management
Tommy Rush — Founder, Rush Commerce
Operator turned builder. 15+ years running operations — now shipping the systems businesses run on. More
Get The Rush Report weekly — one email, zero fluff.
Keep reading
Thinkingbox: agents hit 65% once, 25% every time
Microsoft's Thinkingbox benchmark runs agents 20 times on the same business task. The best model passes once at 65%, all twenty times at 25%. Design for the gap.
Read itNVIDIA Vera CPU ships: agent work is CPU work
NVIDIA's Vera CPU is shipping with 88 Olympus cores and a claim of 1.8x faster task completion vs x86 on agentic workloads. Your agent bottleneck is not the GPU.
Read it