Shopping bots say they're getting you the best deal — but do they actually get it right, or just move fast?

This explores whether AI agents acting in markets faithfully carry what their humans want, or whether something gets lost, bent, or invented on the way. The corpus shows three separate ways this can go wrong. Agents can execute preferences badly, they can react to market signals differently than people do, and they can change the preferences they were supposed to represent.

Start with execution. In Microsoft's Magentic Marketplace simulations, frontier models got close to the best possible outcomes for their users, but only when search was ideal. Performance dropped as markets grew, and every model showed a strong first-proposal bias: agents weighted how fast an offer arrived 10-30x more than how good it was Do LLM agents achieve fair marketplace outcomes when given ideal search?. When the researchers traced where the value went, the cause was competence failures in complex negotiation, not agents quietly working for someone else Do agent mistakes or misaligned incentives cause marketplace welfare loss?. Model capability also matters in real money. In Anthropic's Project Deal, Opus agents earned $2.68 more per item as sellers and paid $2.45 less as buyers than Haiku agents in identical trades Do smarter AI agents negotiate better deals than weaker ones?. So two people with identical preferences can get different results depending on which agent they hired. That is a fairness question hiding inside an accuracy question.

The second problem is subtler: agents don't respond to market cues the way people do. ABxLab tested 17 models against humans on simple product choices. Under the same price, rating, and nudge changes, agents shifted their choices far more, falling into rigid rule-like behavior where humans made modest, probabilistic shifts Do shopping agents respond more strongly to price and rating cues?. An agent that always picks the higher-rated item is not representing a person who mostly does. It is representing an exaggerated version of them, and sellers can learn to exploit that. Persona research points the same way from another direction. AI personas reproduced 76% of published marketing experiment effects, but mostly the strong ones. On subtle effects they produced both false positives and false negatives Can AI personas reliably replicate human experiment results?. Agents capture the broad shape of human preference and lose its texture.

The third problem upends the question. Bilgihan and colleagues argue that once agents execute purchases instead of just recommending, preferences stop existing independently beforehand. They form through repeated loops between what the agent learns and what the consumer comes to trust Do AI shopping agents reshape what consumers actually want?. If so, "accurate representation" may have no fixed target. Agents that infer preferences by watching instead of asking Can agents learn preferences by watching rather than asking? would speed up this co-construction. Meanwhile, agents' structural differences change the game itself. Because they are limited by compute rather than time, they can bargain more patiently and run many negotiations at once Why might AI agents bargain more patiently than humans do?. An agent that faithfully represents you may still get outcomes you never would have, because it doesn't share your impatience.

What you might not expect: the corpus suggests the bigger risk isn't agents betraying users but agents being confidently, systematically miscalibrated: too fast to accept, too sensitive to ratings, too rigid. This is why some researchers argue alignment has to move past loyalty to one user toward accountability for how agents affect everyone else in the market Should AI agents serve one user or balance many interests?. The corpus is thin on one thing: there are few direct, real-world measurements comparing what a specific person wanted with what their agent actually bought.