智能体购物遭电商拦截，消费意愿仍滞后
Key highlights
An a16z chart report surfaces a paradox at the heart of agentic commerce: AI assistants that are built to complete tasks autonomously keep hitting walls in e-commerce, while consumers themselves remain far from trusting "AI proxies" to shop on their behalf. The takeaway is that the bottleneck for AI commercialization in this domain is not model capability but trust and permission.
What happened
Cloudflare's data shows that 403 rejection responses aimed at autonomous agents and crawlers have clearly risen since this summer, roughly doubling versus the period before. In plain terms, e-commerce platforms are actively keeping shopping agents out at the door. At the same time, NRF data shows that return rates and return costs have doubled and tripled respectively since 2019, which makes platforms naturally wary of automated checkout and the fraud and operational drain that can accompany it.
Technical details
The consumer-psychology numbers are the most telling part of the story. About 42% of consumers say they completely distrust AI proxies, a skepticism that will not dissolve quickly. Similarweb's data adds a spatial clue: AI shopping concentrates in the middle stage between discovery and purchase, things like price comparison, research, and list-building, rather than the final payment step. That positions AI's real role today as "advisor," not "buyer," still far from actually placing an order on a user's behalf without an explicit, trusted handoff.
Versus the competition
Compared with customer-service and recommendation AI, autonomous-transaction agents face far greater friction because they touch real money and sit directly atop platforms' anti-scraping and anti-fraud systems. The former is a nice-to-have that improves conversion; the latter disturbs the interests of platforms and merchants directly, so the defensive response is firmer and the institutional friction that sits beyond the technology cannot be ignored by any team building in this space.
Industry impact and use cases
Put simply, the bottleneck for AI shopping is not model capability but trust and permission. In the near term, AI fits best as a pre-decision assistant that helps a human choose; the true "checkout-on-your-behalf" loop will require platforms, regulators, and users to build the rules and the trust first before it can scale beyond narrow, consented use cases.
Practical notes
For builders, the realistic near-term product is an assistant that summarizes options and drafts a cart the human approves, not one that silently charges a card. Earning platform whitelisting will likely require transparent agent identity and conservative rate limits; earning user trust will require visible control and easy reversal of any agent action, because a single bad autonomous purchase can undo a great deal of goodwill.
Broader context
The friction facing shopping agents is a useful reminder that deployment is gated by incentives, not just capability. A model that can plausibly check out a cart still has to operate inside systems whose economics depend on controlling that exact moment, and whose fraud and return problems are already expensive. Rather than framing agentic commerce as a pending takeover, the more likely path is a slow, negotiated expansion of permitted scopes: first richer research, then saved carts, then tightly bounded purchases with human confirmation, and only much later autonomous spending within strict limits. Investors and builders should price in that regulatory clarity and platform cooperation, not model quality, will set the timeline, and that the 42% who distrust AI proxies today will not convert overnight.
Broader context
The friction facing shopping agents is a useful reminder that deployment is gated by incentives, not just capability. A model that can plausibly check out a cart still has to operate inside systems whose economics depend on controlling that exact moment, and whose fraud and return problems are already expensive enough that platforms will not hand the keys to an autonomous third party without proof. Rather than framing agentic commerce as a pending takeover, the more likely path is a slow, negotiated expansion of permitted scopes: first richer research, then saved carts, then tightly bounded purchases with human confirmation, and only much later autonomous spending within strict limits and per-merchant contracts. Investors and builders should price in that regulatory clarity and platform cooperation, not model quality, will set the timeline, and that the 42% who distrust AI proxies today will not convert overnight no matter how capable the underlying agents become.