# AI savings flow to institutional margins, not to tuition — because instruction is under a third of the cost base, and the cost it removes from teaching it partially re-adds as verification

The load-bearing number: instruction is roughly a quarter to a third of operating spend at four-year institutions (Delta Cost Project/NCES; Unverified on exact share), with compliance, facilities, student services, and administration making up the rest. AI automates the automatable share — grading, lecture prep, tutoring, advising — but cannot touch the regulated base, which is set by accreditation, law, and physical plant. So a residential campus that fully substitutes AI for instructional labor cuts total costs by only a sliver. The savings surface as operating margin at endowment- and state-backed institutions, not as price relief at tuition-dependent ones. The tuition floor is stickier than the AI-cost narrative implies.

Second mechanism: the cost AI removes from *teaching* it partially re-adds as *certifying*. When AI can produce student work, the degree's value depends on verifying that the human learned something — assessment redesign, proctoring, oral examination, outcome attestation. This verification function is labor-intensive and grows exactly as teaching labor shrinks, absorbing part of the nominal