Real Option AI: Reversibility, Silence, and the Release Ladder

Abstract

We model the cadence of AI product releases, i.e. quiet spells, reversible patches, and rarer pivots, as optimal exercise of strategic real options under reputational learning. A privately observed technical state follows a diffusion. The firm controls two upgrade options with asymmetric costs and reversibility (a cheap patch and a costly pivot) and a publication-frequency clock, a Cox process whose intensity governs when noisy public performance and safety signals are disclosed. For sufficiently low clock costs the optimal policy posts observable clock-off windows around knife-edge regions. These windows shut down the martingale part of public beliefs, eliminate knife-edge mixing, and collapse behavior to a two-rung release ladder with endogenous triggers, jump targets, and no interior mixing. Within stationary Markov strategies we show that this ladder is uniquely characterized by a boundary-value system with value matching and smooth pasting at triggers and target optimality at jump targets. We endogenize market or platform adoption as a threshold rule in public beliefs and show that leverage creates an irreversibility wedge: the gap between first-best and levered surplus is bounded by the takeover switching cost of the least reversible rung. Patches are debt-insensitive; pivots can be distorted, but only up to that bound. The framework predicts telemetry signatures in firm-authored disclosures: a pre-release cadence dip in publication intensity and intra-month dispersion as the clock is shut off before a major reset; two post-release plateaus in disclosed performance, consistent with patch versus pivot jump targets; and debt-insensitive patch timing in high-reversibility regimes, with leverage effects concentrated in pivots. Unlike option-implied volatility spikes, these patterns reflect the firm's own throttling of technical signals rather than market pricing of event risk.