The Human Layer Economics is out (DOI: 10.5281/zenodo.20096569). It answers the question every serious reader of Paper I eventually asks: if the evidence favors amplifying humans over replacing them, why does investment overwhelmingly fund replacement?
The answer is not that investors haven't read the studies. It is that the incentive architecture makes replacement individually rational while its costs are externalized: pushed to a later quarter, a different budget line, often a different team entirely. Cutting the human layer shows up immediately as saved salary. What it destroys (context, judgment, institutional memory, the ability to notice when the automated thing is confidently wrong) shows up later, diffusely, and rarely with a causal label attached. The paper names this pattern organizational automation bias. A cost you can defer is, to a quarterly system, a cost that does not exist.
The interesting part is not the diagnosis; it is that the deferral window is closing. Three forces are collapsing the gap between when the human layer is removed and when its removal is paid for. Regulatory enforcement is becoming real rather than theoretical. Insurance and liability pricing are beginning to incorporate AI operational risk, which converts diffuse future costs into a present premium. And institutional market access (procurement, audits, certifications) is emerging as a gate that undercapitalized governance simply fails.
History has run this experiment before. SOX, GDPR, ISO 27001: in each case, governance infrastructure was cheapest for the organizations that built it before enforcement concentrated the market, and expensive, rushed and retrofitted for everyone who waited. We see no reason AI governance breaks the pattern.
For Timer this paper is the commercial thesis in academic clothing, and we would rather state that plainly than pretend otherwise: we are building for the organizations that would rather be early than compliant-in-a-hurry.
The series concludes next month with the question all four papers have been walking toward: when everything else in your AI stack is rented, what do you actually own?