So Meta just announced Muse, and I need you to understand why this one is different from the roughly nine thousand other "AI shopping assistant" announcements we've collectively ignored this year.

Muse is not a shopping bot. I realise every company says that about their shopping bot, but in this case it's actually true, which is arguably worse. Muse is a general-purpose personal agent — it sends your emails, books your travel, fills in forms, argues with your phone company about that mysterious charge, and somewhere between planning your dinner party and sending the invitations, it buys the groceries. The purchasing happens as a byproduct of sorting out your life, not as the main event.

This is the bit that should make sellers uncomfortable: the buyer never shows up as a buyer. They show up as someone who asked their agent to handle dinner, and somewhere inside that delegation a transaction occurred on your site, executed by software that has zero interest in your hero banner, your urgency countdown, or your thoughtfully sequenced upsell flow. (It also won't be guilted by the abandoned cart email, but we'll get there.)

Look, "agentic" has been stretched so thin it now covers basically anything with a text field and a loading spinner. So let's be specific about what makes Muse structurally different from the chatbots we've been politely pretending are revolutionary.

First, it acts rather than answers. Previous assistants respond to your prompt and then sit there like an obedient puppy waiting for the next one. Muse takes a goal, decomposes it into steps, and carries those steps out — including steps that unfold over hours or days. Meta describes it continuing to work after you close the app and coming back only when something changes or when it needs your approval before sending a message or completing a purchase. (The approval loop is doing double duty here: it's a safeguard, but it's also a quiet admission that the reliability isn't yet where it needs to be for fully unsupervised transactions. The human stays in the loop because the agent can't quite be trusted to fly solo yet.)

Second, the infrastructure is genuinely novel. Muse runs on what Meta calls a Secure VM — its own virtual computer in the cloud with its own browser. This isn't a cosmetic detail. An agent with its own browser can do everything a person does on a website: open pages, fill in fields, work through checkout flows designed for human hands. It doesn't need your permission or your API integration. If your storefront works for a person, it works for this agent, whether you ever agreed to participate in the agentic commerce era or not. (Consent, as it turns out, is optional.)

Third — and this is the one with teeth — it has memory and initiative. Muse remembers your household's preferences, your usual brands, your dietary restrictions, your recurring needs. It can reorder or substitute without being asked. An agent that knows you always buy the same dog food and can just... handle that... is not a search box. It's a standing intermediary between your customer and every brand that customer used to choose directly.

Here's where Muse cuts against something this newsletter has been tracking for months.

The dominant pattern of the past year has been what we've been calling the discovery-conversion split. AI interfaces have been eating the top of the funnel — the researching, comparing, shortlisting — while actual checkout has stubbornly stayed on the merchant's own site, where saved cards and years of trust already live. The clearest evidence was OpenAI's quiet retreat from in-chat checkout after Walmart reported that purchases completed inside ChatGPT converted at roughly a third of the rate of shoppers sent to its own site. Conversion, it turned out, was stickier to the merchant than the agent-builders wanted it to be.

Muse is trying to reverse this. It completes purchases through Link, the payment layer Stripe built specifically for agents, which generates a single-use card so the merchant never sees the real one. The agent holds the whole funnel: discovery, decision, and conversion in one continuous action.

But here's the important nuance (and I promise it's worth the extra sentence). The earlier attempts tried to pull conversion onto the AI platform's surface, and consumers resisted because that surface felt thinner and less trustworthy than the shop they knew. Muse doesn't ask the shopper to check out somewhere unfamiliar. It checks out on the shopper's behalf, from the shopper's side, carrying the shopper's protections with it. The funnel isn't being moved to the platform. It's being collapsed into the buyer's agent.

For sellers, a reunified funnel that lives on the buyer's side is a stranger problem than a discovery layer you can optimise for. You can learn to feed structured data to a discovery engine. It's considerably harder to sell to a buyer who has delegated the entire act of buying to software that arrives, transacts, and leaves without ever behaving like a customer you can influence.

Strip away the product theatre and several concrete shifts follow.

Your conversion playbook is now addressing an empty room. Two decades of conversion optimisation — the urgency banners, the scarcity cues, the abandoned-cart nudges, the visual merchandising — were all built to move a human mind towards a decision. An agent completing a delegated instruction is indifferent to nearly all of it. It is not tempted by the bundle. It does not respond to the countdown timer. It will not be talked into the larger size. (Your "Only 3 left in stock!" label is doing absolutely nothing for software that can check your inventory API.)

Negotiation is coming. Meta describes the agent as able to negotiate on a person's behalf. Today that's aimed mostly at bills and services, but the direction is unmistakable. There's a useful data point from Project Deal, an internal experiment Anthropic ran, where agents were set to buy from and sell to one another with humans removed. The finding that stuck: the capability of the underlying model mattered more than the instructions operators wrote on top of it. If that holds at scale, the leverage in a negotiation shifts toward whoever has the better model, and the seller facing a well-represented buyer is negotiating against an asset they don't control.

The customer becomes a ghost. When the purchase is completed by an agent on a Meta-controlled machine, paid for with a single-use card, the seller receives an order and very little else. No reliable identity, no email to retain against, no clean read on the journey that produced the sale. Your analytics show a conversion with no legible customer attached to it. The lifetime-value maths that underwrites a lot of paid acquisition assumes you can recognise the buyer again. Agent-mediated purchasing quietly removes that assumption.

The terms of the deal shift toward the buyer. Link brings consumer protections — free coverage for damaged or lost items, price-drop adjustments, fee-free returns. Whatever the fine print, the effect is that a third party sets more of the deal in the buyer's favour, and the seller absorbs more of the returns and adjustment risk as a condition of being bought from at all.

Loyalty becomes a question of defaults. An agent that remembers your usual brands and reorders without being asked turns repeat purchasing into a default-setting problem rather than a brand-loyalty question. The brand no longer wins the reorder by being remembered by a person. It wins by being the agent's default — and the criteria for what becomes a default are set by the agent's maker, not by the shopper's affection for the product.

And underneath all of this: Muse is arriving inside WhatsApp. Not a separate app people need to download. A messaging service that a very large share of the world already opens every day. If agentic shopping has been slow because it lived in destinations people had to choose to visit, embedding it inside the app they already live in is the most credible attempt yet to change that.

None of the above matters if people don't hand the agent the keys, and here the company faces a difficulty its own history created.

Muse only works if a person connects it to email, calendar, payments and whatever else runs their life. That's a larger act of trust than any product the company has previously asked for. Meta has responded with an unusually detailed account of its safeguards — the isolated machine, the separate approval process, the promise that this data stays out of the advertising system.

Security researchers will need time to test those claims. Meanwhile, the structural question is worth naming carefully: Meta's business is advertising. An agent that decides what a household buys could, in principle, rank one product over another, which is influence over purchasing far more direct than an ad in a feed. The company says Muse's data doesn't feed its ad systems today. But if agent-mediated purchasing becomes significant, the pressure to convert the agent's recommendations into a monetised placement will be considerable, and the point at which a helpful suggestion becomes a paid position is exactly the point where sellers would need to pay to be chosen by the software their customers delegated to.

It's too early to know whether Muse works well enough — or is trusted enough — to matter. But the model it represents is the one worth thinking about now, because it inverts the assumption most ecommerce strategy still rests on: that the buyer is a person you can find, persuade, and keep.

Three questions worth holding onto. If the buyer is increasingly an agent, does the discipline of persuasion become a discipline of legibility — and what does "optimising for an agent" even look like in practice? If negotiation becomes routine, how does a brand hold a price against a buyer that never pays list? And if the agent becomes the layer where purchasing decisions are actually made, who pays to be the default — and does that turn the personal agent into the most valuable advertising surface ever built, while insisting (for now) that it's nothing of the kind?

The way I see it, the most unsettling thing about Muse isn't what it can do today. It's that every improvement makes your checkout page a little more irrelevant — and your customer a little more invisible.

👤 Neiro AI - Generate video AI avatars with human-like features, micro-expressions, and emotions

💰 Payback - GPT-4 Vision enabled expense tracker

📖 Capitol AI - AI content that blends storytelling, design & research

🎨 Recraft - Create and edit graphics in a uniform brand style

😍It was great, all I needed

🙂Good, not great

🙄It sucked