Meta built a functional commerce surface in exactly 15 days. Between September 8 and September 23, the company integrated Stripe Link, Shopify, PayPal, and major retail, travel, and grocery partners into the Muse agent. This engineering velocity marks a significant inflection for agentic commerce. Yet as the surface area for these agents expands, the underlying friction remains stubbornly static.
The timeline was aggressive. Following the September 8 launch, Meta moved through a macOS app integration, the Connectors platform, and the number one position on the Apple App Store by September 20. By September 23, the company was announcing retail partnerships with Best Buy, Walmart, and others at its annual Connect keynote, alongside a dedicated hardware device called Muse Charm. Mark Zuckerberg framed the monetization strategy clearly: “We will profit by taking a small fee from transactions.” The model is simple, provided the transactions actually occur.
The problem is that the liability framework has not moved to match this technical speed. When Amazon blocked Muse on September 21, it did so by invoking merchant sovereignty. Amazon argued that third-party applications making purchases on behalf of customers “should operate openly and respect service provider decisions about whether or not to participate.” They cited concerns over Muse failing to identify itself while browsing and the risks of capturing customer credentials. Meta countered that Muse uses a Secure VM architecture with no visibility into passwords or payment details. But this is a technical solution to a legal and trust-based problem.
Three competing approaches to this liability vacuum are now visible. Amazon’s merchant-sovereignty block treats agents as unauthorized intruders — merchants decide whether to participate, and agents that don’t comply get walled off. Meta’s technical safety architecture relies on Secure VMs, a separate Sentinel agent, consumer permissions, and audit trails to mitigate risk at the system level. The existing payment network zero-liability model — where chargeback protections travel with the card credential — offers the closest thing to a consumer safety net, but it was not designed for autonomous transactions.
Data from the PYMNTS/Visa GDSI Merchant Edition highlights the depth of this disconnect. Ninety-three percent of merchants believe the AI provider should bear the financial loss for incorrect purchases. Consequently, only 28% are willing to expose their full inventory to agents, and only 11% of small-to-medium businesses currently qualify as agent-ready. If merchants do not trust the agent, they restrict the catalog — and that caps the potential for Meta’s transaction-fee model.
This brings us back to the Febreze Moment. Despite the rapid expansion of the commerce surface, trust and reward metrics have not shifted. PYMNTS Intelligence data from August 2026 shows that 60% of AI-assisted purchases still close on Amazon. Consumers may use AI for discovery — 23% now start their retail research there, representing 61 million people — but they return to the established merchant for the final transaction. The commerce stack has effectively split: the AI handles the research, but the merchant retains the buy.
OpenAI’s Instant Checkout in ChatGPT, abandoned in early 2026 after low conversion rates, serves as a reminder that technical integration does not guarantee consumer behavior change. Only about 30 Shopify merchants ever went live. Walmart said in-chat purchases converted at one-third the rate of those sent to its own site. The surface area has grown; the conversion funnel has not.
What remains missing is a unified liability framework. Initiatives like the Know Your Agent framework — involving Visa, Mastercard, and Ant International — are in development, but they are not yet operational. Until such a framework reconciles merchant sovereignty with AI-provider financial responsibility, the merchant readiness gap will persist.
Meta has proven it can build the pipes in 15 days. It cannot force the water to flow if merchants refuse to turn the valve. The question is not whether the agent can perform the purchase. It is who pays when the agent gets it wrong.