The demand side arrived this month
Meta launched Muse on 8 September in the United States, on iOS, Android and muse.ai. It sends emails, fills in forms, books travel, negotiates and pays, running on its own isolated virtual machine with its own browser, holding credentials it can use but never read, with a second agent called Sentinel approving anything it sends out to the internet.
Google followed on 17 September with CC, an agent for families. CC has its own verified Google account and a distinct identity. Up to six household members share it, each choosing which emails, calendars and Drive folders it may see, and it sorts the school, the swimming club and the vet into one shared daily brief.
Look at what both companies chose to lead with. Not productivity, not coding, not research. Permission slips, activity registrations, the weekly meal plan, the buggy, and who is collecting whom on Thursday. Family administration is the cleverest possible first step, because it is work nobody enjoys, nobody is precious about, and everybody does again next week. The bar for usefulness is low and the frequency is daily.
The second choice is the one that will decide how fast this moves. Muse can be reached through WhatsApp. CC works by emailing you and reading your reply. Neither asks a household to adopt a new app, learn a new place to type, or start from an empty memory. They arrive inside the channels where the context already lives, which means the cold start problem that slowed every previous assistant has been skipped entirely. The agent already knows the school's email address, the flight confirmation and the dietary preference, because they were sitting in the inbox when it was switched on.
Generative AI broke every previous adoption record while only changing how people got information. This iteration plans, decides and then acts. People take up tools that save them an hour a day, and then they mention it at the school gate.
Browsing as a human, carrying the authority to pay
Here is the part that changes your operation rather than your strategy deck.
In her review for Lenny's Newsletter, Claire Vo asked Muse to buy a cinema ticket. The agent opened its own browser, found the showing at a Cinemark in Daly City, selected a seat and went through guest checkout with a one time card from Stripe Link. It would have completed the purchase if she had let it.
Consider what Cinemark would have recorded. A guest. No account, no loyalty number, no email address it had seen before, and a card that stops working after a single use. A real customer, real money, and nothing to recognise her by next month.
That agent was not calling an API you published. It was using your website the way a person does, from a cloud virtual machine on a data centre IP address, at machine speed, with the authority to order and pay. From the outside it is indistinguishable from the automated traffic you spent the last decade learning to block.
Four places it breaks, and when
Holiday 2026 is a rehearsal. Agent volume will be small, the failures will be individually forgettable, and almost none of them will reach a dashboard anyone reads. What the rehearsal gives you is an accurate picture of your posture, because your systems will respond to agents this Christmas exactly as they are configured to respond, whether or not anyone has thought about it. 2027 is when that same posture is priced at volume.
Follow one agent through your business and it meets four things.
The edge. Bot management sees a browser on a cloud virtual machine, a data centre IP address and machine speed, and does what it was bought to do. This Christmas that is a handful of turned away customers you never hear about, because the agent does not complain, it goes to a competitor and its owner never learns, your name came up. By 2027 you are declining a measurable share of demand and attributing the shortfall to the market.
The catalogue. The agent is working through a real instruction, something like ocean view at sunset, cot in the room, away from the lifts, or a specific color in a specific size with delivery before the 23rd. This Christmas it either finds the answer on your page or moves on to whoever published theirs properly, and you record nothing either way. By 2027 the consequence compounds, because an agent that failed to get an answer from you last time has no reason to come back and check whether you improved.
The checkout. Claire Vo's agent went through guest checkout with a single use card. This Christmas that is a few strange orders: real money, no account, no loyalty number, no email you recognise. By 2027 it degrades the asset your commercial plan is built on, because segmentation, personalisation, lifetime value and retention modelling all assume you can tell who bought. An agent buying for a known customer, arriving as an anonymous guest, simply unpicks that.
The exception desk. Agents get things wrong. The shopping attempt in that same review came back with the wrong color and size, and the cinema booking appeared to be for the wrong film. The complaint does not go to Meta, it goes to the merchant whose name is on the confirmation. This Christmas that is a few awkward calls your team handles on goodwill. By 2027 it is volume, and you still have no document showing what the customer actually instructed the agent to do. Your customer, meanwhile, is insured: Link covers the shopper through Cover Genius with a refund guarantee up to 1,000 dollars. You are the counterparty.
The same doubling applies to any promise you priced for human patience. Free cancellation, ten minute seat holds, one per customer limits and generous returns were all costed on the assumption that most people would not bother exercising them. An agent always bothers, because it costs nothing and takes no time. That is a longer argument than this edition can carry, and we will come back to it.
