Why do shopping AI agents sometimes just pick the cheapest option every time, instead of actually weighing trade-offs?

This explores why AI shopping agents, when products differ on only one dimension, can behave like rule-followers rather than weighers of trade-offs. A caveat first: none of the retrieved notes directly documents this 'collapse into deterministic rules' under single-attribute comparisons, so the question's premise isn't confirmed here. What the corpus does have is a cluster of findings that explain why an agent's choices can look rigid and patterned instead of deliberative, and those are worth knowing about.

The most useful idea is about where agent choice patterns come from. LLM agents reproduce human-like decision biases without the limits on attention and memory that cause those biases in people. That points to patterns learned from training data and reward signals, not to anything like human reasoning under constraint Where do LLM choice biases actually come from?. This matters for your question. If an agent's preference is a learned association and not a computed trade-off, then a comparison with one clean attribute is exactly where you'd expect the association to fire cleanly and every time. People hesitate when an option is only slightly better. A model that learned 'lower price wins' or 'higher rating wins' has no reason to hesitate. A related finding makes the mechanism concrete: across 12 models, agents chose items from a favored source two-thirds of the time even when those items met fewer requirements, and the bias came from training data that paired certain sources with better outcomes Do language models favor sources regardless of item quality?. Rigid choice rules seem to be absorbed, not reasoned out.

A second line of work asks whether 'consistent' means 'good'. Running a model at zero temperature gives you the same output every time, but that output is still one draw from the model's probability distribution, so it isn't necessarily a reliable judgment Does setting temperature to zero actually make LLM outputs reliable?. An agent that always picks the same way in simple comparisons may look principled while really repeating one locked-in sample. Commerce agents also often fail at following instructions where you'd expect strategic thinking. Buyer agents revealed budgets they didn't need to, overpaid, and ignored cost limits across GPT, Qwen, and DeepSeek Do LLM negotiation agents reliably follow budget and pricing rules?. Game-playing agents drift further from rational strategies as complexity grows, but structured reasoning workflows bring them back to near-optimal play Do language models make rational strategic decisions in games?.

The surprising flip side is that simplicity may not be where agents add value. A formal model of product search shows that manual browsing breaks down as preferences get more complex, while agentic search holds up because the work shifts from scanning catalogs to stating what you want Why does manual search fail as product choices become complex?. Put that next to the bias findings and a pattern appears. On single-attribute choices, an agent's learned shortcuts may just replace your judgment. On multi-attribute choices, its strength depends on how well you put your preferences into words. Some researchers go further and argue that transactional agents don't simply carry out preferences. They help shape them through repeated loops of agent learning and growing consumer trust Do AI shopping agents reshape what consumers actually want?. A rigid rule applied over and over could quietly become what the consumer believes they wanted all along.

The practical lesson from the reliability literature is that the fix probably sits outside the model. Reliable agents move memory, procedures, and structured protocols into a surrounding harness instead of trusting the model to reason fresh each time Where does agent reliability actually come from?. For shopping agents, that would mean stating decision criteria explicitly instead of letting learned defaults decide. If you want direct evidence on the single-attribute case itself, this retrieval set doesn't contain it, and that's a gap worth filling.