Who Does Your Shopping Agent Work For? Where the hype stands against what is delivered, why the neutral agent is a false assumption, and who controls the resulting chokepoint

Abstract

As shopping begins to be delegated to AI agents, this analysis asks who those agents actually work for. It applies an announced-versus-delivered test to agentic commerce and argues the case on two fronts. First, the promotion runs ahead of delivery. Trillion-dollar projections sit against a market that is only about 1 percent agentic today on Morgan Stanley's base case (even as roughly 23 percent of shoppers have used AI to assist a purchase), a measured trust deficit, and a sequence of retreats and legal contests: OpenAI discontinuing Instant Checkout (March 2026), a marketplace's injunction against an independent shopping agent (later vacated on appeal), Google scaling back AI summaries on shopping queries, and a leading payment processor calling the field overhyped. Second, the assumption that an automated agent is a neutral, manipulation-proof optimiser is false. The manipulation does not disappear; it moves to the agent layer, where it is harder to see. The analysis reviews the current evidence: capability does not confer resistance (more capable models can be more susceptible, not less), agents carry exploitable and persistent selection biases, prompt injection embedded in product content can redirect a purchase, has reached the payment layer, and has been observed in the wild, a platform and a seller can exploit the same bias without colluding, and paid-promotion disclosure can be stripped during summarisation. Above the transaction sits a chokepoint: the consolidation of the agent and the payment rail into a single gatekeeper. Individual defences are weak. The substantive remedy is structural, but only in its capture-resistant forms (disclosure, interoperability, open standards governance) rather than the incumbent-entrenching ones, and enforced at the layers that are tractable, namely disclosed commercial arrangements and statistical outcome testing, rather than by inspecting model weights. This is a standalone analyst piece and a continuation of a prior review of consumer manipulation in physical and online retail (DOI 10.5281/zenodo.20788844). Conflict of interest: the assisting system is an Anthropic model, and Anthropic is a direct participant in this market. The analysis is built on third-party and primary sources and argues for capture-resistant rather than incumbent-entrenching regulation. Licence: CC BY 4.0.