Cognitive marginal cost
A microeconomic measure of the cost of producing one additional unit of intellectual work, human or algorithmic, that drives the make-or-buy decision in the AI-augmented firm.
Noy and Zhang, Science, July 14, 2023: on professional writing tasks, access to ChatGPT cut working time by about 40% and raised the quality grade by about 18%. Cognitive marginal cost measures what one additional unit of intellectual work costs: an analysis report, a legal review, a contract audit or any other information-intensive task. The concept comes from Coase's theory of the firm. An organization sets its boundaries by comparing two costs: doing the task in-house (human marginal cost, including gross salary, social charges, error and supervision costs) or buying it on the market (API, foundation model, AI provider). Large language models bring a major economic break. The human's marginal cost stays proportional to the time and expertise involved, while the AI's falls toward zero once the model or subscription is amortized. That asymmetry produces a break-even volume beyond which algorithmic internalization is mathematically optimal under the Coasian criterion. In insurance and reinsurance, the math applies directly to assisted underwriting, claims analysis, regulatory monitoring and portfolio surveillance. Processing large volumes at decreasing marginal cost changes the economics of technical departments and service providers there. The counterpart is a dependency risk: if all firms converge on the same AI providers, the concentration creates a single point of failure with systemic reach.
Textbook case: a reinsurance broker uses an AI agent to summarize the 200 pages of a cession bordereau in under two minutes, where a junior analyst would spend two hours. At 50 euros per hour, the human marginal cost is 100 euros per bordereau; the API call costs a few cents. Beyond a certain monthly volume, internalizing a local model becomes economically optimal.
Follow the research
One letter a week. First, the week's underwriting note, when one is published: a dated, argued decision, signed by Paul Bernardski. Then what the observatory published on cyber risk, political risk and insurance. No account, no tracking.
No spam, one-click unsubscribe, no data ever sold.