Every engineer knows Amdahl's law. Almost nobody applies it to AI forecasts.
The rule is simple: your speedup is capped by the part you did not speed up. Make one stage infinitely fast and the system still runs at the pace of everything around it.
Most predictions about AI skip that step entirely. They find a task a model now does well, multiply, and describe the result as organisational transformation. But a 10x improvement in one operation produces almost nothing at the system level if that operation was not the constraint.
So the useful question is not what the model can do. It is what the limiting step actually is.
Sometimes it really is the knowledge work — in which case capability translates into throughput quickly, and the sceptics look silly. More often it is something the model does not touch: approval that takes four days, a compliance review, a customer who signs quarterly, a physical install, a data migration nobody has funded, one person who has to understand the change before it can ship.
This cuts both ways, and that is what makes it useful rather than cynical. It explains why some pilots produce almost nothing after impressive demos. It also explains why a few teams get outsized results from unremarkable models — they happened to automate the actual bottleneck.
Before forecasting from a capability, find the step the work is currently waiting on. If the model does not touch that step, the forecast is arithmetic about the wrong number.