Automation is waiting on liability, not capability

I think the meaningful parts of most white-collar work can already be automated: not “do your job exactly the way you do it today,” but “produce a comparably useful output.” There’s no real technological blocker left. What’s missing is the scaffolding around the model, and that’s a much smaller problem than the one everyone’s still treating this as.

The bottleneck is legal and insurance infrastructure, not robots

I expect software automation, not robots, to boom before physical automation does, and the trigger will be insurers and courts converging on a standard practice for AI-caused error: who’s liable, how it gets underwritten, what “reasonable care” means when the thing making the decision is a model.

Right now companies that could automate a process don’t, because nobody has a decent answer for what happens when the model is wrong. That’s a legal and insurance problem, not a technical one. The day there’s a boring, insurable, litigated-enough answer to that question, the appetite for automation jumps, a lot faster than the underlying models will have improved in the meantime.

Society hasn’t recalibrated its error tolerance for this kind of mistake

People’s risk tolerance is calibrated to the kinds of errors humans make. We don’t have that calibration yet for the kinds of errors LLMs make, so we default to treating it as all-or-nothing: the model has to be right essentially all the time, or it’s not trustworthy at all.

In practice that’s not how anyone actually behaves. People already use tools that are right something like 98% of the time and call that fine — contraceptives sit around that efficacy range and are used everywhere, despite the stakes of getting it wrong being about as high as stakes get. Nobody demands 100% from them first.

LLMs aren’t stably at 98% yet, and even once they are, society hasn’t agreed that “right most of the time, wrong in a specific and knowable way” is an acceptable trade — even though that’s already true of every human doing the same job. Humans aren’t right 100% of the time either, and human error is far less consistent: everyone’s biases are different, everyone fails differently. A model’s failure modes are comparatively stable and predictable per version, which should make them easier to build guardrails around. That’s not how the risk calculus is being done yet, though.

Neither blocker is technical

Put together: the scaffolding to get useful output out of current models already exists. What doesn’t exist yet is a legal and insurance framework that makes deploying it a defensible business decision, plus a shared, honest sense of what error rate is actually acceptable for a given task, and it isn’t “zero.” Once those catch up, I’d expect the visible pace of automation to look much faster than the pace of model improvement would suggest, because the thing that was actually rate-limiting it wasn’t the model.