The capability overhang inside firms

Stripe spends about 1% of its operating costs on AI tokens, and John Collison expects that to rise to 5%, 10% and 20%. He said so to Sam Altman on the Cheeky Pint podcast. Altman said the gap between what most people use AI for and what the most productive users get out of it keeps widening, and that more companies use too little AI than too much.

Altman used the term capability overhang for models that labs hold back from release. I think it fits the gap inside firms just as well.

In payments and finance operations, I think a lot of the value from AI over the next few years will come from models that already exist today. Better models will keep coming, and areas like scientific research still have plenty of room for big gains. For the payments and finance processes I look at, though, the extra gain from each new model seems to be levelling off.

My own tests on payments and finance tasks back this up. The best models, Claude Opus in particular, already handle most cases well, and on routine steps many models score about the same. So on the tasks I've tested, a better model would add little.

Hence the challenge for many firms now is catching up with that overhang. For finance and operations teams, that means redesigning processes around AI to get the return that's already available.