Autonomous shopping agents need math more than language, according to a new study on arXiv (2607.04708).
Researchers tested three policy types for AI buying agents deciding when to pull the trigger on a purchase within a fixed window: a stationary regime assuming prices follow a Poisson process with a known distribution, a Bayesian regime that updates on uncertainty about that distribution, and a robust regime that assumes only price bounds and optimizes for worst-case outcomes. All three were evaluated against 48,933 timestamped price observations across 367 Amazon products drawn from the Keepa price-tracking database. The stationary and Bayesian policies held their own on average consumer surplus, but the robust policy was the clear winner at the 10th percentile — the scenario where timing goes badly and protecting the buyer matters most.
That finding cuts against the current trend of dropping large language models directly into agentic purchase flows and hoping the model figures it out. The paper's evidence points in the opposite direction: LLMs are better at selecting which regime to apply and calibrating the inputs than at making the actual buy-or-wait call. In other words, the model is a wrapper around the math, not a replacement for it.
Shops like Amazon have already built dynamic pricing systems that exploit exactly the kind of uncertainty these policies are designed to handle — so the adversarial framing of the robust regime is less academic than it sounds.