Fermi Estimates for AI Risk

AI is usually the last line in the risk register still scored in colors, and a color cannot be added up, weighed against a control's cost, or shown to be wrong. This technical white paper sets out the alternative: a Fermi estimate that prices one way an AI system can cause a loss as a short chain of factors, each pinned to a company record, a measured sample or a named source, and multiplied out to the loss to expect in a year.

What the paper covers

Part 1, what a Fermi estimate is. The origin of the method at Trinity and in the piano-tuner problem, why a chain of checkable factors beats a single guess, what a 5×5 matrix cannot do, and the anatomy of an AI risk chain.

Part 2, the method. The four rules that keep the number defensible, the four roles that produce an estimate and why they are separated, and three AI systems of a fictional machinery manufacturer priced end to end with arithmetic you can check, then summed against a budget and re-run as evidence arrives.

Part 3, at scale. Why hundreds of systems built on models that change every few months need the estimate re-run on a schedule, how the four roles run as agents on the Modulos platform against the company's own records, how the number becomes a curve, and how a configuration change is priced before anyone flips it.

The paper closes with where the method stops and a sources table checked against primary texts.

Who it is for

The people who produce the number or have to trust it: risk managers, CISOs and AI governance leads, controllers and internal audit, and the advisers who brief boards. It assumes intelligence and no prior knowledge of Fermi estimation.

About the figures

The worked example is constructed and labeled illustrative on every exhibit; the manufacturer is fictional and its figures are invented, in EUR. Every external claim carries a named source, and the platform facts were verified against the product at print date.