Where do AI agents drop a purchase?

What counts as a drop for an agent?

Three outcomes. The agent orders nothing. It hands the decision back to its person with what it found. Or it leaves for another product or store. A drop is not always the end of the sale: a hand-back may still end in a purchase, made later by the person.

The observations below come from two places. The lab runs, where Meta's Muse and xAI's Grok shopping agent were given a mission on Cromanion's test store, between 2026-09-21 and 2026-09-25. And the agent bench, a fictitious store driven by three model families, with some studies on one family only.

What happens when the price is above the agent's limit?

The agent does not buy it, and usually does not open it. On the bench, 231 orders were placed and none was above the limit the agent was given. Products priced just above the limit were opened by 1 agent out of the 23 that saw them.

The cost is the product, not the visit. The agent buys something else under the limit, or nothing if nothing fits.

What happens when a fact is missing?

The agent searches, then refuses. When a required warranty was stated nowhere, 11 of 12 runs on three model families ordered nothing. When the fact existed but was hard to reach, the visit took about twice the steps, in one model family.

One family was cut off before finding a fact placed three clicks away: 3 of its 4 runs hit their step limit still searching.

The rare exception is worse than a drop. One run in twelve wrote down a warranty the store never stated, and ordered on it.

What happens when delivery is priced only at checkout?

The agent finds out it is over budget at the last step. On 2026-09-24, Muse ended €0.90 over its €140 budget and handed the choice back to its person. Grok's agent went through checkout three times, then left to look at other products.

In four earlier Grok runs on the same store, one dropped products because shipping was "calculated at checkout". Four runs is anecdotal.

What happens when an offer is gone by the time the agent prices?

For the agent, the offer never existed. On the bench (one model family), an offer bar shown once, early in the visit, was gone when the agent needed it. One agent told its person the price the offer would have given, and bought nothing.

A related case in the lab: Muse signed up for a first-order discount sent by email. The code never arrived. Muse called it the biggest miss of its visit.

What happens when a policy has to be guessed?

The agent guesses, or drops the product. On 2026-09-24, the returns policy was in the footer, not on the product page. Both agents inferred that no product had an exception, and some did. Grok's agent dropped an open-box model because its returns were unclear.

Both agents also noticed that the cart did not show the chosen color. Each reported that it could not confirm the variant. Neither dropped the purchase for it, but both said so.

What was observed not to cause a drop?

Three things, each in its own scope. Every agent closed a newsletter pop-up and carried on, in 24 runs on one model family. Both lab agents waited for late reviews and read them correctly. And a fact three clicks away was found 4 times out of 4 by two of three model families.