Agentic Commerce

What It Is

Agentic commerce is the pattern where an AI agent runs the buying loop on a user’s behalf instead of just suggesting what the user should buy. The user states an intent in natural language, and the agent handles product discovery across a catalog, applies the right loyalty discounts and stored payment method, completes checkout, books the logistics, and manages returns or complaints if anything goes wrong. The human signs off on the intent, not on every click.

The category got its first true production deployment this week when Alibaba wired its Qwen AI app into Taobao and Tmall with Alipay-native checkout attached, reaching roughly 300 million monthly active users with access to a 4-billion-item catalog. The launch demo had Qwen ordering 40 cups of bubble tea, applying loyalty discounts, and completing checkout from inside a chat thread without the user ever opening the Taobao app. That is the difference between an AI agent that recommends a purchase and an agent that completes one.

Agentic commerce sits next to agentic-coding as the second consumer-visible category where the agent does the work, not the human. Coding agents close pull requests. Commerce agents close orders.

How It Actually Works

Three layers have to line up for an agent to actually buy something. First, the agent needs catalog access, structured enough that it can match a fuzzy intent (“forty bubble teas, two flavors, delivered by 4pm”) to specific SKUs with current inventory and pricing. Second, the agent needs identity and payment plumbing, meaning it has to know who the buyer is, which payment method to charge, and which shipping address to use, all under permissions the user has explicitly granted. Third, the agent needs after-sales hooks, the same APIs a human would use to check order status, file a complaint, or trigger a return.

The Alibaba deployment works because Taobao, Tmall, and Alipay are all owned by the same company, which means the catalog, the payment rail, and the logistics stack all sit behind one set of credentials. Western implementations are messier because the catalog usually lives at one company, the payment rail at another, and the fulfillment at a third. That is why Microsoft’s Copilot Cowork ships partner connectors instead of native checkout, and why most US agentic-commerce demos stop at “added to cart” instead of “ordered.”

The agent itself almost always sits behind a managed-model API like Claude or its peers, with the commerce-specific tool-use wired through structured function calls into the catalog, payment, and logistics services.

Why It Matters Right Now

Agentic commerce reframes who the e-commerce site is selling to. For twenty years the buyer was a human scrolling product images, and the entire optimization stack, search ranking, recommendation widgets, ad placement, photography, was tuned for that human’s eyes. When the buyer becomes an agent, none of that matters. The agent reads structured product data, not photos. It compares prices through an API, not a price-comparison site. It cares about return rates and ship times, not the lifestyle photo on the product page.

That shift is going to break the consumer ad business and the SEO business at the same time, and it will reshape how merchants list products. The first signal to watch is whether agents start surfacing different SKUs than the top organic results, because that gap is the early evidence that the optimization stack has the wrong target.

How TWO Uses It

TWO covers agentic commerce as the consumer-facing twin of the agent infrastructure stories that dominate the macro section every week. When Anthropic signs a $200 billion compute deal, that is interesting to operators but invisible to readers. When Alibaba ships an agent that orders bubble tea inside a chat, the same readers can see the future arriving on a screen they already use. Scott’s editorial filter is whether a story changes what a non-technical professional will touch this quarter, and agentic commerce passes that filter every time it ships at scale.

The operator decision Scott has been weighing is when to start instrumenting his own commerce surfaces, the Wise Operator’s Starter Kit checkout in particular, for agent traffic. The current best read is that agent traffic will look like API traffic before it looks like browser traffic, which means the work to do this year is making sure structured product data is clean, prices are queryable, and return policies are machine-readable. The browser-facing checkout is a separate problem that does not need to be rebuilt yet. The signal that the timing is changing is when a checkout starts seeing 5%+ of orders coming from a non-browser user-agent, because that is the point an agent has begun completing transactions on real customers’ behalf.

What to Watch Next

Three signals will tell you agentic commerce has crossed from demo to default. First, when a major US retailer (Amazon, Walmart, Target) ships its own commerce agent inside its own app rather than waiting for an OpenAI or Anthropic plugin. Second, when Visa, Mastercard, or Stripe ship a “delegated payment” credential built specifically for agents, separate from the existing card-on-file primitive. Third, when the first ad buyer publicly cuts spend on a category because agents are skipping the ad layer entirely.

Until those three hit, agentic commerce remains a category-defining demo with one production deployment in China. After they hit, it becomes the default expectation for how online buying works.