Caterpillar will spend $100M training 118,000 employees on AI and autonomy
The five-year programme covers field repair, digital twins, mining and legacy code. The company's own framing is that the models are not the hard part.
Caterpillar is spending $100 million over five years to train 118,000 employees on AI and autonomous systems, covering field repair, digital twins, mining operations and legacy code maintenance.
The per-head figure is about $850 a year. That is a serious training budget for an industrial workforce and a rounding error next to what the same company spends on autonomy hardware — which is the point the programme is built around.
The claim that separates this from most corporate AI announcements
Caterpillar's stated position is that deployment depends on experienced operators redesigning workflows, not on installing models.
That is the opposite of how most enterprise AI has been sold, and it matches what the evidence keeps showing. Meta's push towards an "AI-native" workforce reportedly produced a 220 percent increase in code changes and a 36 percent rise in delivered features — alongside a 40 percent increase in incidents. More output, more breakage, because the constraint was never typing speed.
Meta has since discontinued its AI-adoption dashboards and token-consumption scoring in performance reviews, following legal challenges. Measuring AI usage turned out to measure usage.
Caterpillar is measuring something else: whether the person who knows how the mine works can redesign how the mine works.
Why legacy code is on the list
The inclusion of legacy code maintenance among the four target areas is the most revealing detail, because it is where industrial companies actually hurt.
Caterpillar runs embedded control software written over decades, in languages whose expert practitioners are retiring, on machines that cost millions and cannot be experimentally rebooted. The knowledge is in the heads of people who will not be there in ten years, and there is no market solution — you cannot hire for it.
That is a well-shaped problem for a model: large, poorly documented, highly structured, and desperately in need of anyone who can read it. It is also a problem where a wrong answer damages equipment, so the engineer supervising matters more than the tool.
The workforce framing others are getting wrong
Compare UBS, which the Financial Times reported this month will require graduate and intern candidates to demonstrate AI proficiency for 2027 junior positions.
UBS is filtering its intake for a skill. Caterpillar is upgrading the people who already know the domain. For work where the scarce knowledge is about the equipment rather than the software, the second is the more defensible bet — a graduate who is fluent with models and has never been in a mine is not the constraint on mining autonomy.
Both approaches assume AI proficiency is a durable skill worth selecting or training for. That is not obvious. The tools have changed substantially every six months, and much of what constituted proficiency in 2024 is now handled by the product.
What does not go out of date is knowing which problems in a business are worth pointing a system at, and that is domain knowledge.
What is not measurable yet
Caterpillar has not published what outcomes the programme is meant to produce, how it will be measured, or what happens if the training does not translate into deployments.
A $100 million training commitment with no stated success metric is a bet that broad literacy pays off diffusely. That may well be right. It is also unfalsifiable, which is convenient for whoever proposed it.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters