the side effect was the company

a founder told me this week that the most valuable thing his company owns is something he never set out to build.

he was trying to make a system that runs without being prompted. to get there he had to invent infrastructure that did not exist, because the way everybody currently wires these things up, an agent goes back and reads the raw material again from scratch on every single run. so he built a layer that reads everything once, understands it, and hands the agent the answer.

his metaphor stuck with me. a librarian who has read every book, rather than a catalogue that tells you which shelf to walk to.

the side effect is that running the same work costs a tiny fraction of what it cost before. that side effect is now the business.

the number underneath it is the part people are not ready for. somebody put the cost of agent tokens at a couple hundred dollars per employee per week. run that across a company of forty thousand people and you are looking at a line item that would make a cfo cancel the whole programme, which is roughly what has been happening. most enterprise ai deployments are quietly failing, and the postmortems keep blaming the model when the bill is what actually killed it.

here is the flip. everyone assumes inference gets cheaper forever, so cost stops mattering, so nobody should build here. he argues the opposite. as inference gets cheap, cheapness stops being a moat, and the value moves up the stack to whoever holds the understanding rather than the raw data. the pipes get commoditised. the wisdom does not.

and the model providers eventually have shareholders. cheap forever is a belief, not a plan.

but the thing i actually want you to take is the shape of how this happened. he did not go looking for the valuable thing. he went after a hard goal, hit a wall that had no product behind it, built the missing piece himself because he had no choice, and the missing piece turned out to matter more than the goal.

that is how a lot of real companies start, and it is why i distrust the founder who has the whole map before the walk. you cannot find the wall from the desk. you find it by trying to do something specific and difficult and running out of road.

the discipline is noticing. most people build the workaround, ship the original thing, and never look back at the tool they made to survive the week.

what did you have to build along the way, and is it better than what you were building?