AI agents raise average order value by changing what shoppers find, not by pushing harder at checkout. The published production numbers: shoppers who engage Total Wine's AI agent convert at four times the rate of a traditional session and spend about 20% more per order when they buy. The scoping travels with the claim every time it is used: engaged shoppers versus traditional sessions, never a sitewide average. This post explains the mechanism behind those numbers, where it works, and why it looks nothing like the recommendation widgets e-commerce has been bolting on for fifteen years.
What do the production numbers actually say?
Total Wine & More, the largest independent fine-wine retailer in the US (294 superstores, roughly 8,000 wines per store), runs a Zowie AI agent that both resolves customer service and sells. Among shoppers who engage it, conversion runs at four times the rate of a traditional session, and order value runs about 20% higher when they buy. At Decathlon, conversion from support interactions to purchases rose 8%, and support-driven revenue rose 20%. Wojciech Ćwik, Decathlon's Omnichannel Project Manager, stated it directly: "Thanks to Zowie, our conversion rate from support interactions to purchases grew by 8%."
Read the scoping honestly and the numbers get more interesting, not less: they measure what a conversation does to a buying decision, which is exactly the variable a retailer can now choose to deploy.
Why does a conversation change what people buy?
Because most of a catalogue is invisible to most shoppers. Search and navigation only surface what a customer already knows to ask for; everything else depends on merchandising real estate that only a few products can occupy. A conversation has no shelf. When a shopper describes a need in their own words ("a gift for someone who likes bold reds but under fifty dollars"), the AI agent can answer from the entire catalogue and current stock, including products the shopper did not know existed.
Cole Lillie, Total Wine's Director of Digital Product Management, described the surprise in the production data: "I knew it could answer a question for a customer. I didn't think it would change what they buy." Customers were buying products they had not known about, because they learned about them in conversation. That is the discovery mechanism: the AOV lift is a merchandising effect, not a persuasion effect.
How is this different from a recommendation widget?
A recommendation widget guesses from behavior; a conversation asks. Widgets correlate ("customers who viewed this also viewed"), which works at the margin and fails exactly where order value lives: considered purchases, gifts, technical products, and anything where the shopper's need does not match their click history. A conversation gets the constraint directly from the customer (occasion, budget, recipient, experience level) and can do what a good floor associate does: explain why this one, offer the step-up option with a reason, and complete the basket with what actually belongs together. In-store, your customers get an expert; online, they get a search bar. The AI agent closes that gap, which is why the discovery effect shows up first in categories with deep catalogues and knowledge-heavy purchases.
Where does conversational selling work best?
The pattern from production: deep catalogues where discovery is genuinely hard (Total Wine's 8,000 wines per store is the extreme case), considered and knowledge-heavy purchases where shoppers want guidance (sports equipment, electronics, supplements, apparel fit), gifting, where the buyer is by definition outside their own expertise, and replenishment businesses, where the conversation can carry proactive reorders, cart recovery, and win-back on channels customers already use. The common thread is that the shopper has a describable need and the catalogue has a better answer than the shopper's search query would find. Where the purchase is pure habit at a known price, the selling upside is smaller and the service upside carries the case.
Why does selling come after service?
Every production deployment worth citing follows the same sequence: the AI agent earns trust by resolving customer service to a measured standard, then selling switches on. The logic is architectural as much as political. An AI agent that cannot be trusted with a refund cannot be trusted with a sale; service is where it proves that discounts, return eligibility, and policies run deterministically (the model never decides money) and that quality holds when every conversation is evaluated. Total Wine's sequencing is the reference: 64% resolution first, then the selling numbers. For the CFO, the same sequencing means the business case compounds on one platform: cost per conversation falls on the service side while conversion and order value rise on the selling side. The full service picture is in AI customer service for e-commerce.
How to measure AOV impact honestly
Three rules keep the measurement believable. Scope the comparison: engaged-shopper conversion and order value versus traditional sessions, stated exactly that way, never blended into a sitewide claim the deployment did not make. Watch the counterfactual: discovery-driven AOV lift should come with flat or rising CSAT and no return-rate spike; a lift built on pressure would show up in returns. And attribute conservatively: support-attributed revenue (Decathlon's +20%) and conversion among engaged shoppers are the two cleanest cuts, because both compare a conversation cohort against its own baseline. If a vendor quotes an unscoped sitewide revenue claim, ask for the cohort definition before you believe anything else in the deck.
Watch the discovery mechanism on a real catalogue at getzowie.com/commerce, or read the Total Wine case study, where the numbers come from.
