Organisations deploying AI across knowledge work often cannot tell whether it is working. That is a failure of measurement infrastructure that long predates the technology.