The average enterprise spends 2,068 $ per employee per year on AI. The median spends under 200 $. An average with ten times its median describes nobody, so this positions you instead: your industry, your headcount, and the one split no published source carries.
AI spend per employee varies by a factor of five between sectors, because knowledge work has more automatable surface per head than capital-intensive work. Pick yours.
The third input is the one that matters and the one nobody publishes: how much of your AI work produces something that leaves the building. An internal error costs rework. An external error becomes a contract, a filing or a claim, and redoing the work does not undo it.
Adoption in production is published for four departments and not for the other three. The three without a figure carry no bar, because a cell we have not measured is worth less than an empty one. Each row is per 100 people in that department, at the intensity and governance gap you set above, so nothing here assumes how your headcount is distributed.
Read the order. The three most adopted departments are customer service, IT operations and marketing, and two of the three are external: adoption ran ahead precisely where a wrong output cannot be recalled. HR is the uncomfortable one, internal and heavily adopted for screening and regulated, which is the combination that produces a claim rather than a rework.
This instrument apportions your spend. It deliberately does not multiply external exposure by a risk factor, because no honest source publishes one and inventing it would make every other number on this page suspect. What a wrong external output actually costs in your operation depends on your cycle, your gates and who signs. Four weeks measures it. See the audit.
Four weeks, one team, one real decision cycle. You leave with your own figure instead of an industry average, and with the twin of your own house.