How will AI agents progressively take over online shopping?

A Boston Consulting Group report frames the e-commerce journey as four steps — trigger, discovery, comparison, purchase — and argues that AI agents will disrupt each step through four "evolutions" toward full autonomy. In Evolution 1, consumers "trigger the process and embark on discovery and comparison within a large language model (LLM) or AI chat interface such as ChatGPT," then click through to the merchant's own product or checkout page. In Evolution 2, exemplified by "AI-powered answer engine Perplexity Pro," consumers select a product and check out from within the AI interface itself, using a wallet or stored payment method, "never visiting the merchant's website." In Evolution 3, the agent conducts "each stage from discovery and comparison through purchase with minimal to no human input," selecting the product "based on historic preferences and preset criteria" and checking out with a stored payment method. In Evolution 4, the agent owns the entire journey and takes on the purchase step itself, "intelligently optimizing the consumer's payment methods based on offers and rewards to maximize consumer benefit."

BCG's reasoning for staging the disruption this way is that capability and adoption do not move together: it expects "the evolutions to overlap as adoption varies by product category, ticket size, merchant, and other factors." The stages track a steady transfer of control — first over discovery and comparison (Evolution 1), then over the checkout surface itself (Evolution 2), then over product selection (Evolution 3), and finally over how the consumer pays (Evolution 4) — rather than a single threshold a market crosses all at once.

The staging adds a temporal dimension to How do advisory and transactional AI agents differ in function?: Evolution 1 is squarely advisory (the agent informs, a human still transacts on the merchant site), while Evolutions 2 through 4 move progressively deeper into the transactional category that typology defines, with Evolution 4 adding a capability — autonomous payment-method optimization — that a simple advisory/transactional split does not distinguish. Evolution 3's description of an agent acting "based on historic preferences and preset criteria" echoes the mechanism in Why does manual search fail as product choices become complex?, where agentic search is modeled as substituting preference articulation for manual comparison; BCG's framework describes the same substitution narratively rather than formally. The consumer-adoption figures BCG reports for this same study are in Who wins and loses as AI agents reshape shopping?.

The four evolutions are a forward-looking typology, not a measured adoption curve: the report gives no data on how much current agentic shopping activity falls into each stage, no timeline for progression between stages, and no quantification of what "minimal to no human input" means in Evolution 3. The framework should be read as BCG's scenario vocabulary for talking to payments-industry clients about where to position, rather than as evidence that markets are moving through these stages in this order or at any particular pace.