Meta tests human concierge workers behind its Muse AI agent, sparking debate over the future of agentic commerce

Meta's Human-in-the-Loop Agent Signals a Training Data Play, Not a Product Permanence

Meta is testing human concierge workers to handle phone calls placed by its Muse AI agent, according to Bloomberg reporting. The setup—contractors quietly completing tasks the digital agent can't yet execute—looks like a crutch but is actually a data collection mechanism disguised as customer support.

The constraint is real: Muse can summarize calendars and handle routine tasks, but it falters on complex, real-time negotiations. Ordering flowers or resolving a customer service issue requires judgment the current model lacks. Rather than wait for voice models to improve, Meta is stuffing humans into the loop to generate training data for the next iteration. As voice capability advances, the human step compresses or vanishes. This is not a long-term business model—it's a bootstrap strategy.

Why this matters for product adoption

The human fallback obscures a deeper problem: usage may be the real bottleneck. Consumer agents live in the same friction-free digital environment as Netflix and TikTok, but they demand intentionality. Even users with personal assistants in the physical world skip them on weekends to "chill." Time-on-app metrics suggest users engage episodically—they get excited, perform tasks, then disappear. Weekly active user comparisons (OpenAI's chosen metric, not daily) reflect this pattern. An agent's power is irrelevant if people don't open it.

The commerce battle is about infrastructure and data capture

Meta isn't alone in the field. Shopify has partnered with Muse, but the integration only works with Shop Pay, turning agent integration into a lever to move merchants onto Shopify's payment platform. The risk is clear: Walmart's ChatGPT integration saw conversion rates drop to one-third of the website's native rate, with smaller cart sizes. Consumers ask agents for specific items; they don't discovery shop.

Amazon, meanwhile, has structural leverage. Their e-commerce business generated $36 billion in net income in the last twelve months, but advertising revenue was $72 billion—ad revenue is now twice the e-commerce profit. Without ads, Amazon's core retail business is unprofitable. This means Amazon will fight agentic commerce harder than most. They already rejected OpenAI's instant checkout integration in favor of a multi-billion-dollar ad deal. Now they're blocking browser automation to prevent agents from ordering through them.

The workaround game is already real: developers propose running custom agents on home WiFi, wrapping browser control through alternative browsers to evade Amazon's detection. It's cat-and-mouse, not a solution.

Where value gets captured

Shopify's bet assumes merchant fragmentation can compete with Amazon's logistics. If an agent can't find socks on Amazon, it shops Shopify instead—and five independent Shopify merchants distribute inventory across different regions. The agent picks whichever can deliver fastest tomorrow. It's not seamless like Amazon, but it works at scale. This requires Shopify merchants to stock breadth beyond the single-product D2C model they've traditionally built.

Sponsored agentic flows emerge as a revenue play. An agent notices you like Marvel films and have a Friday free; it proposes tickets to the new Spider-Man film, books with friends whose outings sync to your calendar, and connects them via text. The brand pays for the placement. Fandango, the theater chain, and the film studio all benefit, but sponsorship splits are unclear. The long tail of small business will test this at scale.

The open web gambit

Meta, Stripe, and Shopify are framing this as a return to an "open web" that competes with Amazon. The irony is dark: Meta killed the open web in the last cycle by walling off Instagram and Facebook, making everything closed and ad-driven. Now they're selling open-ness as a virtue. The difference is that this time, if agents are the distribution layer, being the fallback option (rather than the sole destination) creates a structural advantage for platforms with good API access.

Walmart, despite acquiring Vibe, appears to be missing the opportunity. Without a lucrative ad business to defend, Walmart could lean hard into agent-first commerce and position itself as the default option for agentic purchasing. Instead, they're copying Amazon's ad play—and losing the chance to be category different.

The knife fight between Amazon and everyone else hinges on whether agent-driven commerce can fragment across multiple merchants, or whether logistics and advertising moat will keep Amazon dominant. Agents will route to whichever merchant delivers fastest or cheapest. If Amazon loses enough marginal transactions to Shopify merchants and other platforms, the moat cracks. If Amazon moves faster and integrates agent commerce deeper than competitors, the moat holds.

Human workers in Muse's backend are just the opening move.