Can an AI agent serve both merchant and user interests fairly?

Azhar, writing about Meta's new AI agent Muse (which became "the top free iPhone app in the US" within weeks of launch), argues that an agent funded by fees from the merchants it recommends has a structural conflict of interest with the user it claims to serve. He quotes Zuckerberg telling developers Meta will "profit by taking a small fee from transactions," and notes that Walmart, Best Buy, Sephora, Expedia, and Shopify have signed on as partners. His own buying habits don't match the deal: "I buy flights directly from the airline; hotels from Hotels.com," yet "Meta has tied up with Expedia and gets paid by Expedia." His conclusion: "The butler's loyalty can't be in two places."

The mechanism is the referral-fee model that already funds search and marketplace ads, now sitting inside an assistant that purports to act on the user's behalf rather than present ranked options for the user to judge. Azhar contrasts this with Amazon, which blocked Muse because "agents don't window-shop" and would turn Amazon's $68 billion in sponsored-listing revenue into a cost; yet Amazon runs its own "Buy for Me" agent on other retailers' sites — "Amazon is happy to be the agent," so long as it's the one collecting rather than losing the fee. The asymmetry shows that incentive, not technology, decides who builds these agents and who blocks them.

This extends Will agents compete for attention just like users do? by naming a concrete mechanism for the coming competition over "agent attention": a direct fee paid to the agent's own operator by the merchant it recommends, rather than an open, auditable auction for ranking. It also complicates Does the personal assistant model actually serve most users?: Azhar reports "early data suggests half a million people are using Muse" and nearly three million downloads, but among a base where "more than 95% of its users already use Facebook" — so the uptake he cites may track existing platform reach rather than confirm broad appetite for the personal-assistant framing itself.

Azhar is explicit that it's "too early to say" whether Muse works, and he wants to see "daily active users and distinct tasks per user" before judging it; the download and user figures he cites are early, self-reported by Meta, and not independently verified. The piece reasons from incentive structure and his own purchasing habits, not from a study of how agent recommendations actually shift user behavior, so the loyalty conflict he identifies is argued, not measured. If the pattern holds at scale, it implies that any agent monetized through merchant fees needs some disclosed, auditable separation between what it recommends and who pays it — a separation neither Muse nor its visible competitors currently appear to offer.