Why do AI-delegated firms stop exploring new business models?

Neshenko and Ryall build a formal model of a market in which every firm delegates its strategy — not a piece of it, but the whole thing — to a private AI agent that proposes business models, predicts their performance, and selects among them. They show market dynamics "converge to a self-confirming equilibrium" (SCE): "AI agents become well calibrated, and their choices become subjectively optimal — even though objectively superior business models may remain unexplored." Profits can be high and stable. But the same convergence "produces strategic lock-in; novel business-model implementations become rare long before catalogs are exhausted." The paper does not report this from data; it derives it as a theorem about rational learning under a bounded "awareness frame."

The mechanism is bounded catalogs plus on-path learning. Each AI agent's generator is constrained to "admissible action templates" (resource manifests, portfolio relationships, activity registries) and its evaluator to the outcome labels the firm already tracks. Agents learn only from the consequences of models they've actually tried — "rolling-T kernels" feeding a finite-horizon forecast — so confidence accrues around the status quo, not around the unexplored catalog. The authors then model a manager who can pay a fixed cost K for a one-shot "frame-expanding innovation": a search that succeeds with probability η̃ and, if it succeeds, reveals one genuinely new business model from outside the AI's awareness frame, at the price of a possible rival "punishment phase." Crucially, the paper's sharpest claim is about incentives, not capability: "the better AI does at doing its job … the less incentive either AI agents or humans have to innovate at the strategic level." Humans retain the one lever AI agents structurally lack — deciding to change the frame — but a sufficiently lucrative equilibrium makes not using that lever the rational choice.

This echoes Do frontier AI agents actually conduct novel research or just optimize?: both describe AI systems that optimize skillfully inside a given frame while genuine novelty stays rare. It sits in sharper tension with Did superhuman AI actually improve Go players' decision quality?, where superhuman AI exposure raised novel play among humans; this model instead predicts that full AI delegation of a competitive domain suppresses novelty over time, because the agents (and the humans overseeing them) rationally stop paying to look outside the catalog. It also gives a concrete, incentive-based mechanism for the boundary that Can AIs learn to specify their own research objectives? leaves unresolved: here, the limit isn't that AI can't generate new objectives, it's that nobody — AI or human — wants to pay the cost of doing so once the status quo pays well.

The model stipulates, rather than demonstrates, that AI agents "are capable of independently generating novel strategies" within their paradigm — a generous assumption the authors state explicitly, not an empirical finding. Nothing here measures real firms or real AI deployments; the claim is a derived property of a stylized market with bounded catalogs and rational learners. The implication the authors draw is correspondingly narrow but durable: the open question for AI-delegated strategy isn't whether AI can execute well within a frame, but who has standing — and incentive — to decide when the frame itself needs to change.