NextFin News — China’s peak shopping festival still opens with the usual sprint—brands crossing large sales marks within hours—and with a new detour. The same household staple can cost less after a chat assistant applies channel coupons than it does inside a short-video storefront. A gap of roughly ten yuan is enough to move a casual order. Artificial intelligence is no longer only answering product questions. It sits on the path to checkout, carrying e-commerce’s oldest lever: the subsidy.

Merchants are on another timeline. Industry versions of office AI already report a meaningful share of paying customers from online retail. Detail pages, campaign assets, and performance reviews were budget lines before any model learned to write a caption. Tools that compress those lines can invoice against money that was going to be spent anyway. Inventing a durable consumer habit is hard. Taxing an existing operating budget is a business plan that needs less persuasion.

The voucher moves with the door

A year ago, platform “help me choose” features still sent people back into the core shopping app to finish the job. This season, conversation products can recommend live rooms, assemble a basket, and—in some stacks—complete payment without a deliberate return to the old home screen. Whoever receives the half-formed need first gains a chance to decide which shelf appears next.

That influence does not require the assistant to become the merchant of record. Overseas measurements of visits that reach retailer sites via AI referrals have shown higher conversion and higher revenue per visit than non-AI sources. Domestic competition is intensifying for the same structural reason. Payment and fulfillment may remain inside Alibaba or JD. The ranking of options may already have been shaped elsewhere. For a general assistant tied to a short-video commerce graph, steering the order is a claim on future merchant spend. For an incumbent retailer, owning checkout is no longer the same as owning the decision.

Subsidies make the experiment rational for users who do not yet trust the model’s taste. A chat path that is simply cheaper trains a reflex: open the assistant before opening the store. The platform pays for the habit; merchants and live rooms still clear the transaction. Figures labeled “AI-driven sales” therefore need a split. Many tickets are redirected demand, not invented demand. Redirecting an order is still worth a fight. It is not proof that dialogue created a purchase that would never have happened.

A good guide can be a weak salesperson

Consumer research shows wide trial and narrower satisfaction. Large majorities report having tried AI shopping features; far smaller shares say the tools match products well. Familiar complaints recur: weak respect for hard price limits, mixes of cheap and expensive items that force manual cleanup, conversations that run long without removing work. When a budget must be restated after the model “helps,” the session has failed as labor-saving software.

Better models can reduce clumsy mismatches. They cannot erase a conflict built into the product. Shoppers want fewer mistakes and less waste. Platforms want conversion. Merchants want volume. Faithful advice might be that the appliance at home is adequate, or that a cheaper generic is enough—good counsel, bad retail. Classic search hid that tension behind pages of results. A single conversational answer owns the choice and invites a blunt question: whose side is this voice on? Turning novelty into trust means accepting sessions that end in smaller baskets or no basket. Merchant software rarely faces that outcome. Consumer assistants provided by commerce companies live inside it.

Selling the meter, not only the map

Alongside guided shopping, major marketplaces have begun listing model usage itself—monthly and annual packs from foundation-model vendors sold through ordinary storefronts. Early category growth rates look dramatic from a small base; absolute volumes in flagship shops remain modest. Platforms have been restrained about heavy subsidies here. The contest is where an already-paying API customer settles the bill: the vendor’s site, or a familiar e-commerce checkout with comparison and renewal attached.

That channel fight rhymes with the shopping-assistant fight. The metrics should not be blended. Token sales measure distribution of AI services. They say little about whether an assistant understands a household’s trade-offs, and less about whether AI is transforming retail demand as a whole.

Production gets cheaper; attention does not

Merchant AI sells against a clearer job. Listings must ship. Tests must run. Creative must refresh. Launch materials describe open-to-operate cycles cut from weeks to days and research-and-mockup spend collapsing from tens of thousands of yuan into token fees. Even when a new product fails, the work product had a price. Vendors can collect before the first unit sells. Deeper links into seller analytics and advertising platforms keep the same customer inside one operating system after the pictures are done.

Efficiency does not make competition gentle. When every store can generate many creative variants before lunch, attention stays scarce. Money saved on production migrates into bidding for impressions and clicks. Platforms can earn twice: once on the productivity subscription, again on the media that becomes more necessary as content supply explodes. “AI cuts merchant costs” is an incomplete sentence. Who captures the surplus—tool vendor, traffic landlord, or retailer facing equally accelerated rivals—decides whether the technology softens the treadmill or speeds it up.

This festival season, the clearest AI invoices replace work merchants already paid humans and agencies to do. Consumer-facing assistants still buy trial with coupons and still argue, implicitly, about loyalty. One side is collecting on workflows. The other is still renting curiosity. Both are real businesses. Only one, so far, looks like a habit that can pay without a subsidy attached.