May 12, 2026
Where AI pays for itself in a forty-person company
I keep a list of the pilots that worked. All of them are boring.
It comes up in every first meeting now, usually about forty minutes in, usually phrased as a worry rather than a question. I have started answering with a list I keep in the back of my notebook, because the honest answer is unglamorous and it helps to have evidence.
The pilots that worked in companies this size have three shapes. Drafting: the first version of a quote, a job advert, a reply to an awkward customer, written by a machine and finished by the person who would otherwise have stared at a blank page for twenty minutes. Sorting: inbound mail or applications tagged and routed so someone reads the right forty rather than all four hundred. Reading: a long supplier contract or a year of survey answers, summarized well enough to decide whether it is worth reading properly.
What those have in common is that a person still signs the output, and the cost of a bad draft is a rewrite rather than a lost customer. That is the line I hold, and I have not yet met a case at this scale where crossing it was worth the risk.
One finding I did not expect, and did not much like: the saving shows up as slack, not as headcount. Nobody leaves. People simply get their afternoons back. Deciding what those afternoons are for is a harder conversation than the technology one, and it is the one I now try to have first.
MMB