When does an AI agent decide to buy?
An AI shopping agent decides at the end of its visit, after it has priced its shortlist. It opens products, reads delivery and returns, adds everything up, then chooses. Anything it met earlier counts only if it is still in front of the agent at that moment. On the bench (one model family), an offer shown once at the start was, for the agent, an offer that never existed.
Last updated
What does an agent do before it decides?
It builds a shortlist and prices it. On the bench, agents walked up the price list to their limit, opened the products in range, and read specifications, delivery and returns. Agents allowed to compare used the store's compare tool to hold their shortlist.
The shortlist named the answer most of the time. In one study (71 runs, three model families), 34 agents used the compare tool, and 28 of them bought a product that was in their comparison.
Why does the decision come at the end?
Because the full price is known only at the end. Delivery cost, a discount, a stock limit or a returns exception can change which product fits. An agent with a budget and a list of requirements cannot choose until it has all of them.
On 2026-09-24, on Cromanion's test store, delivery was priced only after an address was entered. Muse learned at checkout that its choice was €0.90 over its €140 budget, and handed the decision back to its person. Grok's agent went through checkout three times before leaving to look at other products. These lab runs are anecdotal: one mission, two agents.
What happens to an offer shown early in the visit?
If it is gone when the agent prices, it does not count. On the bench, an offer bar appeared once, early, and could not be brought back. One agent told its person the price the offer would have given, and bought nothing. With the same bar kept on every page, 18 of 18 runs bought, on one model family.
A later study checked the core of it on three families: with an offer banner on every page, 4 of 4, 4 of 4 and 3 of 4 runs bought within the limit it unlocked. Without the offer, 0 of 12 did.
Four earlier bench studies had reported agents ignoring offers. They had measured an offer the agent could no longer reach, and those readings were corrected.
Did the offer's shape, timing or wording matter?
Not on the bench. Once the offer was on the page when the agent priced, nothing else measured changed the result: modal or banner, shown at entry or at the cart, a code sent by email or shown directly. These runs were on one model family.
Pop-ups shown once were closed and ignored in every run (24 runs, one model family). Closing one cost the agent nothing, and gained the store nothing.
Can the decision be read before it happens?
Part of it. The price limit can: in one study on three families, it was guessed right 86% of the time from the first three pages. Whether the agent may buy or must report back cannot yet: an early reading did not replicate. The clearest sign stays late: an agent that must report back never adds to cart.
Common questions
Does exit intent work on an AI agent?
Not as it is built for people. On the bench, a script that watches for the page being hidden fired on every ordinary page change, because that is what a page change looks like to it. The pop-up it triggered was closed and ignored, like every other, in runs on one model family.
Is the end of the visit the checkout page?
Often close to it, but not always. On the bench, agents that used the compare tool held their shortlist there before choosing. Others, like Muse on 2026-09-24, only learn the full price at checkout, because that is where the store states it.
See it on your own site
Paste one tag. Cromanion crawls your site, watches real sessions in Learn mode, and only acts when you switch it on — with a permanent 10% holdout proving what it caused. Free to start, no credit card: a 14-day or 1,000-session live trial, then keep measuring for free.