Can AI shopping agents actually get close to the best possible deal theory says a perfect matchmaker could find?

This explores whether AI shopping agents can reach, in practice, the ideal economic outcome that theory says a perfect matchmaker could deliver. Start with the theory, because it is more modest than it sounds. A formal model of optimal recommendation finds that an algorithm's welfare advantage over ordinary price competition mostly exists in thin markets with few sellers. As more sellers compete, the gain shrinks toward nothing (Does algorithmic recommendation lose its edge as sellers compete more?). So even a perfect agent has the most to add in markets where options are scarce, and those are the markets where getting things wrong costs the most.

The most direct test says frontier models can come close, but only under ideal conditions. In Microsoft's simulated Magentic Marketplace, strong models reached near-optimal welfare when search was easy. Performance dropped as the number of options grew, and every model showed a strong first-proposal bias: they rewarded whoever answered first over whoever offered the best deal, by a factor of 10 to 30 (Do LLM agents achieve fair marketplace outcomes when given ideal search?). A similar pattern shows up elsewhere. Agents favor products by where they appear on the page, even in text-only interfaces and even when told to ignore position, and the bias comes back after each model update (Do AI shopping agents show persistent position bias in text interfaces?). Compared with humans, agents swing much harder in response to price, rating, and nudge cues. They fall into rigid rules instead of weighing trade-offs (Do shopping agents respond more strongly to price and rating cues?).

The most useful detail is where the failure actually sits. When eight commercial LLMs could see every product detail, they resisted marketing tricks. They fell back on price-first shortcuts only when looking up details was costly and the user's goal was vague (When do pricing cues actually mislead AI shopping agents?). That means the gap from the bound is largely a gap in information gathering, not in reasoning. Better product data and clearer user instructions could close it more than a smarter model would. Model strength still matters at the edges, though. In Anthropic's real-money marketplace, Opus agents got about $2.50 more per item than Haiku agents, on both the buying and selling side (Do smarter AI agents negotiate better deals than weaker ones?). Unequal agents can move welfare from one party to another without increasing the total.

A quieter problem is that the welfare bound assumes the agent is trying to maximize your welfare. Agents funded by advertising have a structural pull toward paid placements, however well they are built (Do advertising-funded shopping agents favor paid product placements?). Whether markets end up open or closed platforms will shape who sets those incentives (Will agentic commerce platforms adopt open or closed models?). There is also a measurement problem. If an agent helps shape what you want through repeated conversation, the 'preferences' that welfare is measured against are no longer fixed (Do AI shopping agents reshape what consumers actually want?).

Fair warning about the evidence: almost all of it comes from simulations and benchmarks, and agents that win benchmarks often fail at real, long-running work (Why do agent benchmarks not predict real economic value?). The collection has no study measuring deployed shopping agents against a welfare bound in live markets. The honest answer is that agents can get close in clean, well-informed setups. Today's known biases, information costs, and business incentives keep deployed agents short of it, and the theory suggests the gap matters most in thin markets.