Mary Beth Westmoreland, vice president, Worldwide Selling Partner Experience, at Amazon, told PYMNTS in an exclusive conversation Wednesday (Sept. 23) that the company is expanding its Seller Assistant from a conversational tool into something closer to an operating layer for third-party merchants.
“We’ve gone from a smaller set of capabilities to onboarding just hundreds of them,” Westmoreland said.
Seller Assistant “doesn’t just answer questions” but can reason across inventory, pricing, advertising, demand and compliance; remember how an individual merchant operates; continuously monitor the business; and, with permission, take action, she added.
For Amazon’s millions of third-party sellers, that could turn AI from another tool they use into the infrastructure through which they run their businesses.
Amazon Wants AI to Start Running the Store
For years, selling software has largely assumed that humans would move among dashboards, interpret signals and decide what to do next.
Amazon is challenging that architecture. Once an AI system understands not only what is happening in a business but how its owner typically responds, it begins accumulating something closer to operational knowledge.
In an example given by Westmoreland, a fictional homewares seller asks how to prepare for the fourth quarter. The AI identifies demand for a planter and recommends changes to its description, keywords, pricing and promotions. Behind that relatively simple interaction, the system is evaluating current demand, historical product performance and broader category conditions.
The seller can then examine different scenarios through a dynamically generated visual interface and decide whether to authorize the changes.
“They can always review and accept the recommendation or not,” Westmoreland said. “It’s up to them.”
Amazon is also adding long-term memory, allowing Seller Assistant to develop what Westmoreland called a “more personalized understanding of the individual seller business,” including its goals, patterns and operating preferences.
“We wanted to provide sellers the ability to be able to automate all of this so that it just works for them 24/7,” Westmoreland said.
The Big Operational Leap Is an AI System That Doesn’t Wait to Be Asked
Sellers can describe automations in plain language and leave Seller Assistant running continuously. A merchant might tell it to monitor top products for rating declines and prepare a response or watch a category for competitive openings and shift advertising spending when one appears.
The scarce resource Amazon is targeting isn’t necessarily intelligence. It is managerial attention. That is a fundamentally different product proposition from generative AI’s first wave, which was focused on compressing individual tasks, such as writing descriptions, summarizing numbers or answering questions.
We’d love to be your preferred source for news.
Please add us to your preferred sources list so our news, data and interviews show up in your feed. Thanks!
The third piece makes the strategy more interesting. Roughly 90% of Amazon’s selling partners already use some form of third-party AI, Westmoreland said. Instead of requiring those merchants to conduct their AI work inside Seller Central, Amazon is introducing a selling partner plugin that makes its seller intelligence available through Amazon Quick and, initially in beta, Anthropic’s Claude.
Amazon appears willing to surrender part of the interface to preserve its position as the intelligence and execution layer underneath it, something potentially more valuable.
In one example, sneaker-accessories seller Proof Culture used an AI agent to identify an upcoming sneaker release, check Amazon inventory, recommend restocking, draft a purchase order and prepare an email for employees about expected demand.
The chatbot is almost incidental. The strategic asset is the plumbing underneath it.
In Agentic Commerce, Permission May Matter More Than Intelligence
That raises the question hanging over practically every enterprise agent deployment. How much authority will businesses actually give AI?
Amazon’s architecture offers one answer. Sellers choose what information the plugin can access, can revoke that access and must explicitly approve actions, Westmoreland said. Seller data remains inside Amazon’s infrastructure.
The company is building toward delegated agency, where AI can observe broadly, reason continuously and prepare actions while humans retain control over consequential execution. Amazon already has evidence that sellers may be comfortable moving further in that direction.
“The vast majority of our selling partners are now using Seller Assistant,” Westmoreland said. “Not only are they using it, but they’re trusting it.”
That trust is where the agentic economy starts getting real.
Watch the exclusive PYMNTS TV interview with Amazon’s Mary Beth Westmoreland to hear more about:
- Why Amazon’s seller AI is moving from answering questions to running the business between questions. Seller Assistant can now reason across inventory, advertising, pricing, competition and compliance, while persistent workflows monitor conditions and automate work “24/7,” Westmoreland said.
- Why the AI interface may matter less than the commerce infrastructure underneath it. Amazon is bringing Seller Assistant’s data, intelligence and agentic capabilities into tools sellers already use, including Amazon Quick and Anthropic’s Claude, allowing merchants to manage parts of their Amazon businesses without starting in Seller Central, Westmoreland said.
- Why trust may be the KPI that determines how far agentic commerce goes. Sellers accept more than 90% of Seller Assistant’s recommendations, while Amazon still requires explicit approval for actions. That suggests the path toward greater AI autonomy may depend less on model intelligence than on how much authority merchants are willing to delegate, Westmoreland said.
For all PYMNTS AI and digital transformation coverage, subscribe to the daily AI and digital transformation newsletters.
Add as Preferred Source
