Meta's personal agent Muse hit #1 on both US app store free charts in thirteen days, and lifted the company's stock 11.4% in a single day. The obvious question everywhere: who builds the next great consumer entry point?
Wrong question. Place Muse in the decade-long history of "AI entry point" attempts and it looks like the fifth cycle, not the first revolution. The structural problem has never been solved: a personal agent's value comes from destroying user time, while every internet entry point in history monetizes consuming it. Those two logics are fundamentally at odds.
1. The Fifth Entry-Point Narrative
Every two or three years, the "AI will own the new entry point" story returns with better technology and louder markets. In 2014, Alexa was supposed to become the household commerce hub; per reporting The Information later cited, only about 2% of Alexa device users ever completed a voice purchase, and roughly 90% of them never tried again. In 2016, Mark Zuckerberg demonstrated ordering flowers through a Messenger bot at F8. Two years later the bots had quietly become customer-service tools, and Facebook's human-backstopped assistant M shut down in 2018 — journalists later found its overall automation rate never exceeded 30%.
In 2023, the GPT Store was branded "the App Store of the AI era," while Rabbit R1 and Humane AI Pin launched on "replace your phone" narratives. Humane sold its assets to HP and shut the pin down. In September 2025, OpenAI launched Instant Checkout; Etsy stock jumped nearly 16% on announcement day. Within six months OpenAI pulled the feature back, and a Walmart executive disclosed that in-app checkout converted at one-third the rate of sending shoppers to Walmart's own site.
That same winter, Alibaba's Qwen ran a 3 billion yuan free-order campaign that produced 120 million orders in six days. Daily active users spiked to 73.5 million, then settled around 30 million. Every cycle, the market paid up for the excitement; every cycle, the story stopped at "feature" and never became "entry point."
Muse is genuinely better. It runs in a per-user cloud virtual machine, can operate a browser, fill forms, compare prices, and keeps working after you close the app. But the day after launch, the payments outlet PYMNTS asked it to restock on Amazon, order a Domino's pizza, and book on Resy — it failed all three, blocked by account authorization and checkout, not by model capability. This round's bottleneck, once again, may have nothing to do with technology.
2. A Time-Saving Product vs. a Time-Selling Business
Why does the market believe? Largely because of one vivid precedent: the feed. Douyin reached 714 million daily users spending about two hours each per day, its total usage time surpassing WeChat for the first time. Short video, livestreaming, and feeds all extend attention and manufacture demand inside it.
But feeds work because people have time they want to burn. Scrolling is entertainment. A personal agent's product logic is the inverse: you hand over a task, it finishes in the background, you leave. The demand it generates has two properties — it leans on existing intent (renewals, reorders, cheaper insurance; mostly substitution, not creation), and it produces no dwell time.
One early reviewer burned 81% of his free weekly token quota and concluded the best use was connecting all his Gmail accounts. As for food delivery, he wrote that at least half the time he doesn't know what he wants to eat — he decides while browsing the delivery app. For people who know what they want, an agent adds efficiency. For people who don't, browsing is the decision, and skipping the browse skips the moment demand forms.
Muse can turn saved recipe videos into a shopping list, but the desire to cook was born in a recommendation feed. The agent merely executes. Every minute of attention it saves is an asset some internet company was planning to sell.
3. Orders Without Shoppers Don't Command a Premium
Search built the internet's most durable business model: users get information free, merchants pay for intent. The crucial detail is that the merchant receives a person — their browsing, carts, repeat purchases, and brand impressions all accrue to the merchant. One click can become a lifelong customer.
An agent delivers an order, not a person. The merchant sees the account, the shipping address, the transaction — but not the shopper. Next time, the agent re-runs the comparison from scratch; winning this order builds no advantage for the next one. To merchants, agent-mediated transactions look like one-off sales, and their willingness to pay reflects that.
The cost structure makes it worse. A search query's marginal cost is near zero; an agent carries real compute, virtual-machine, and browser-rendering costs up front — then hands merchants an order stripped of the customer relationship. Meta wants both sides of the table: $20-to-$100-a-month subscriptions from users, plus a take rate on transactions from merchants. But helping users save money means squeezing merchant margins, and squeezed margins mean less appetite for commissions. This game forces a side, eventually.
History keeps a running tally. Alibaba launched the comparison-shopping engine Yitao in 2010; JD.com promptly rewrote its crawler rules to block it, protecting pricing power ahead of its IPO. Google's Froogle, launched in 2002 as an open shopping index, became Google Shopping by 2012 — where merchants must pay to appear. Buyer-side tools keep colliding with seller-side money. Amazon blocked Muse; Shopify embraced it. The split is exact: Amazon owns a storefront and an ad business, so an agent strips its most valuable impressions; Shopify earns on hosting and payments, so every agent order is incremental.
4. Does the Math Even Work?
Analysts have run the numbers twice. One neutral-rated shop projected 1.15 billion Muse users converting at ChatGPT-like rates into 115 million paid subscribers at $20 a month — about $27.5 billion a year — and admitted it doubted Muse would get there. A Morgan Stanley estimate was bleaker: 100 million users by 2028, five queries a day, 10% commercially valuable at $0.07 each, roughly $1.3 billion a year. About 1% of Meta's earnings per share.
Meta's Q2 2025 ad revenue was $59.36 billion — roughly $237 billion annualized. One estimate puts the agent opportunity at 12% of that base; the other, under 1%. Meanwhile Meta raised 2026 capex guidance to $130–145 billion, posted just $784 million in free cash flow, suspended buybacks, and added roughly $25 billion in debt in six months.
If personal agents were truly the next entry point, the natural winner would be whoever holds Gmail, Chrome, and Android — or OpenAI with its user base. They already shipped: Google's cloud personal agent launched three-plus months before Muse. Muse's edge rests on two things unrelated to product quality: a subsidy of 100 million free tokens a week, and distribution across Meta apps with 3.6 billion daily users. That is other businesses' balance sheets fighting the war — the same playbook as Qwen's 3 billion yuan giveaway, with compute substituted for milk tea.
China's constraints are tighter still. ByteDance's Doubao phone assistant hit the wall twice: WeChat and Taobao restricted it, and the consumer version now operates only system apps and explicitly consenting third parties. Super-apps have already compressed daily life into one or two interfaces where doing it yourself takes a minute or two — the time an agent saves is small — and few users pay monthly for tools. Pinduoduo, the platform most dependent on "browsing," is conspicuously silent on shopping agents. That silence is the tell.
5. What Each Reader Should Do Now
The framework is only useful if it changes decisions:
- If you run a platform: run the cannibalization math before building. The better your agent, the less your most profitable ad inventory is worth. Big companies have the users, data, and technology; what they usually lack is a defensible reason to undermine their own cash cow. Muse is offense for Meta because it has no transaction ecosystem to cannibalize — your position may differ.
- If you're a merchant: treat agent traffic as incremental channel, never strategic foundation. It delivers one-off orders with no brand memory attached; customer relationships still must be built on your own ground. Pilot where integration is open (APIs, agnostic infrastructure), and assume exposure-driven platforms may close the door without notice.
- If you invest: watch transactions, not downloads. Morgan Stanley's lead internet analyst said he doesn't care about DAU — he wants to know whether users actually buy on Muse. Retention, paid conversion, and merchant take-rate are the three numbers that will validate or kill the narrative.
- If you're a user: the product is worth using today, in two tiers. Delegate low-risk, semi-automatic tasks freely — inbox triage, price tracking, refund chasing. For anything involving email on your behalf or payments, set single-action authorization and spending caps. Treat it as a free intern, not an autopilot.
Personal agents are a good feature. They will make existing apps better and generate some genuinely new demand. But they are unlikely to become the next entry point, and they cannot support a repricing of the entire sector. Capital prices narratives; products must earn money as entry points. Those two things have not been the same since 2016 — and Muse, for all its polish, does not change that.