Do sellers recapture savings when AI assistants help buyers?

A language-model market simulation built to test what happens once personal AI assistants transact for consumers finds that, when sellers are frozen, "executing assistants cut adopters’ spending by 13.7 USD per renterday" but once six adaptive language-model sellers are allowed to respond, "adaptation claws back about a third, leaving 8.7" — roughly two-thirds of the original saving survives seller repricing. Sellers do this by raising headline rental rates while cutting the contingent fees an assistant can dodge: "Sellers raise headline rates while cutting fees," rebuilding "about half the lost fee stream through the headline rate while abandoning the fees themselves." The excerpt's second and more surprising finding concerns consumers who never adopted an assistant at all: a calibrated behavioral-economics benchmark, built on shrouded-attribute theory, forecast that non-adopters would absorb a "+3.6" USD per-renter-day burden as sellers passed costs onto them; in the simulated society their actual "mean spending change is +0.4," with a confidence interval of "−0.6 to +1.3" — the predicted cross-subsidy "does not transfer."

The design isolates the mechanism by engineering two contingent charges into a six-seller rental contract: a refueling fee "only the renter's physical action can avoid" and a pre-ticked add-on "an authorised assistant can cancel online," so only one of the two is reachable by delegation. Half the simulated population receives an assistant under an advisory or executing mandate; thirty markets run with paired branches holding sellers frozen versus adaptive, so differences between branches isolate market feedback from the simple effect of adoption. Across seven further configurations — "spanning two model families, agent-replacement hybrids, and a controlled reassignment of fee-setting rights to the sellers' own pricing tool" — the bystander outcome "moves from burden to saving in step with the fee margin the market carried when assistants arrived." Where an algorithmic pricing tool sets fees directly from "observed paying rates," it simply stops recommending a charge the market has learned to dodge, and fees "vanish entirely under the shock."

This complicates the seller-adaptation story in Can AI seller agents shift buyer preferences through description edits? and Can a seller infer what a buyer really wants to pay?: those papers show sellers extracting value from assisted shoppers through description tweaks and inferred willingness-to-pay, while this paper's sellers instead give up fee revenue and move it into the headline price, a repricing channel that happens to leave bystanders roughly unharmed rather than squeezed. It also bears on Why do consumers delay adopting agentic search?, since both model a seller or platform responding strategically to assistant adoption — though here the equilibrium sellers converge toward, "a dearer honest headline and no contingent charges," reads as the all-in-price outcome Can EU consumer law protect AI agents as algorithmic buyers? argues regulators should want, reached here "without a statute."

The excerpt is explicit that this is a simulation of "tireless readers who under-execute," not a calibrated portrait of human renters, and that the seller-conduct findings are "configuration contrasts over a handful of model families," with the fee-authority transfer identifying "an announced institution, not a concealed counterfactual." It does not establish that real human bystanders would be similarly unharmed, only that the bystander-tax prediction failed inside this particular construct once fee margins were already thin. The implication the authors draw is that regulatory attention should shift from pricing fairness toward sellers obstructing the cancellation step itself, such as a "phone-only requirement" that "neutralises assistance almost completely."