When an AI shops for you, the hard part stops being 'what's out there' and becomes 'what do I actually want?'

This explores why, when an AI agent searches on your behalf, the bottleneck moves from "finding options" to "saying what you want." The clearest statement comes from a formal model of product search Why does manual search fail as product choices become complex?. In manual search you pay for every product you look at, so as products gain more attributes that matter to you, browsing stops being worth the effort. You settle early or give up. An agent can scan a huge catalog at almost no cost to you, so catalog size stops being what limits you. What limits you now is how well you can turn your preferences into a signal the agent can act on. In the model, the sensible move is to spend your effort refining that signal rather than looking at more options. The scarce resource becomes articulation cost.

This raises a question the question doesn't quite ask: what if people don't know their preferences well enough to state them? Bilgihan and colleagues argue that with agents that actually make purchases, preferences often aren't formed in advance at all Do AI shopping agents reshape what consumers actually want?. They take shape through repeated back-and-forth between what the agent learns and how much the consumer comes to trust it. So "articulation" isn't a single act of describing a fixed want. It's partly done with the agent, and the agent's suggestions help decide what you end up wanting. If articulation is the new scarcity, whoever shapes that conversation has a lot of influence.

One response is to skip articulation and infer preferences by watching behavior. Memory architectures like M3-Agent organize what they observe around specific people and things, separating "what happened" from "what I know about this person," so they can act on preferences nobody stated Can agents learn preferences by watching rather than asking?. The catch: conversational LLMs are passive by design. They answer, but they rarely start a conversation or ask the clarifying question that would help you put a vague want into words Why can't conversational AI agents take the initiative?. So the cost of articulation falls mostly on the user, and current agents do little to lower it.

There is also a twist about who benefits. A related model finds that platforms prefer agentic search before it actually pays off for consumers Why do consumers delay adopting agentic search?. Platforms count only the sales it brings in. Consumers also count the effort it costs them, and that now includes the effort of explaining themselves. That gap means hesitant adopters can be acting rationally: switching only pays once the cost of stating your preferences falls below the browsing effort it saves. On the supply side, search effort is getting cheap and scalable. Agents' answers keep improving as they get more search steps, much as they do with more reasoning steps Does search budget scale like reasoning tokens for answer quality?. This strengthens the main argument: when machine search gets cheaper, human articulation becomes the bottleneck.

The library has one formal model making the core claim and a few nearby threads. It doesn't yet have empirical studies measuring articulation cost in real shopping sessions, so the claim is best read as a strong theoretical argument rather than a settled finding.