Customers can't bring their own agent — seven weeks after blocking Meta's Muse, Amazon opened "text Alexa"; Airbnb said bookings from Muse and Instinct are "very nonmaterial" and promised its own agent for 2027; and Ramp wrote that US business AI spending has stood still for two months

Last week brought three numbers that show how the door through which an agent buys on a customer's behalf opens and closes. On November 6, Amazon let a subset of US customers text its own shopping agent (Alexa for Shopping) to search, fill a cart and complete a purchase — seven weeks after it blocked Meta's agent Muse from ordering on Amazon on September 20. On November 5, Airbnb said in its Q3 results that bookings coming from Muse and Instinct are "very nonmaterial," and that it will launch its own travel agent, not a chatbot, in the first half of 2027. Payments-data company Ramp wrote that the share of US businesses paying for AI reached 55.9% in September, but token spending had stood still for two months. Outside agents are quiet, platforms are building their own, and business spending has reached the top of the S-curve. We lay out how the three numbers are counted and which columns a small team running apps and commerce should fill in this week.

This article was written with AI (Claude) and reviewed by an editor. Our company is a paying customer of Anthropic, the maker of Claude, and of OpenAI, and we run apps, commerce and AI channels ourselves. In item 3 below, Ramp's data shows Anthropic ahead of OpenAI (business adoption 43.8% vs. 40.6%; token-spend share 60.5% vs. 35.1%). Because those numbers could read in favor of a company we use, we note the nature of the sample and its caveats alongside them and add no evaluation on top. Figures announced by companies are marked "as announced by the company," and anything from a single source is marked with that source. Quarterly results and announcements from listed companies (Amazon, Airbnb, Meta) appear, but this is not an investment opinion on any company and not material for investment decisions. The valuation and credit-rating reports that ran alongside the Ramp article are investment material, so we did not use them.

In NO.42 we wrote that "Amazon has two doors": into the front door that is its own search, it lets outside stores in for free; at the front door that is its own store, it keeps outside agents out. This week showed what is happening behind that second door. Amazon put its own agent in the spot it had blocked, and Airbnb put a number on how few bookings outside agents bring and announced its own agent too. This is the third number in the place where NO.35 read "agents placed hundreds of orders on a delivery app" and NO.40 read "four agent ride requests in Uber's busiest week." And alongside it came a number, counted from payments data, on how far business AI spending — the side paying for those agents — has come. We separate facts, interpretation and "if it were us."

1. Amazon — where it blocked Muse, it put "text Alexa" (facts)

On Friday, November 6, Amazon opened shopping by text message with Alexa for Shopping (Amazon News, as announced by the company). Text the number that appears in your phone's contacts as "Alexa Shopping," and the agent Amazon calls a "personal shopping agent" finds products, fills the cart, completes the purchase and sends delivery updates in the same thread. Turning it on takes two steps: accept the terms and link a phone number. The announcement lists these capabilities: send a to-do list with several items in one text and get suggestions as they are found; set up and change weekly grocery orders; pull ingredients from a recipe photo and add them; reorder past purchases with the customer's approval; remember preferences such as "organic only"; find similar products from a photo; track prices and alert you when they drop (up to a year of price history); check order status; and compare products by details, reviews and price. Products from other stores also appear through Shop Direct, which we covered in NO.42, and the conversation carries across text, the shopping app, the web and Alexa devices. It is for a small, invited subset of US customers, with "millions more customers in the coming weeks," and anyone not invited can sign up at amazon.com/text-alexa-shopping. The scale the company gave: "more than 400 million customers used Alexa for Shopping in the last 12 months" and "usage of its agentic features has grown more than eightfold since the start of the year." The announcement does not define either number (whether one use counts, or what "agentic features" covers). Rajiv Mehta, vice president of Conversational Shopping, is quoted: "your personal shopping agent is just a text away." There is not a single line about outside agents.

Something else came before that text message. On September 20, Amazon blocked Meta's personal agent Muse (released September 8) from placing orders on Amazon (SiliconANGLE and PPC Land, two sources). The Amazon popup shown to Muse users read: "Continued access by an unauthorized AI agent violates Amazon's Conditions of Use, to which our customers have agreed." Amazon's reasons come down to two. One is that Muse does not identify itself as an agent while browsing the site (Amazon's terms require agents to put an identifying string in their requests); the other is that Muse appears to enter customers' account pages and purchase history and to store credentials. Amazon's statement to GeekWire: it is "fairly straightforward that third-party applications that offer to make purchases on behalf of customers from other businesses should operate openly and respect service provider decisions about whether or not to participate." Meta's reply was that Muse "has no visibility into people's passwords or payment methods" and that it is "actively working to expand Muse's data protection guardrails." Citing GeekWire, PPC Land noted that Amazon's own agentic purchase feature, Buy for Me, identifies itself and lets brands opt out, and that Amazon has also moved to restrict shopping agents from Google and OpenAI (secondhand).

One court note. In November 2025 Amazon sued Perplexity, maker of the Comet browser with a built-in shopping agent, under anti-hacking law and won a preliminary injunction in March 2026, but on August 4 the Ninth Circuit ruled that "the party accessing Amazon's computers is the user, not the AI company," undoing that win (PPC Land and SiliconANGLE, two sources). Both outlets added that the court "did not impair Amazon's ability to regulate access through its terms of service," and a rehearing petition was denied on September 10. That is why the Muse popup cited the Conditions of Use, not hacking. The same week, TechCrunch reported that DoorDash said last month that third-party agents had placed "hundreds of orders" in total and that it would focus on its own text-based agent, launched in September (TechCrunch, one source; the same remark as the "hundreds" in NO.35).

2. Airbnb — "bookings from Muse and Instinct are very nonmaterial," and its own agent comes in the first half of 2027 (facts)

Airbnb reported Q3 results on November 5. The numbers on the company newsroom (as announced by the company): revenue $4.9 billion (+19%), gross booking value $27.7 billion (+21%), nights and seats booked +16%, net income $1.9 billion. The same post has an agent paragraph. "We're building our own AI agent to help guests plan and book trips. We plan to start rolling it out to guests in the first half of 2027." It is not a chatbot. "We don't think a traditional chatbot is the right way to plan a trip," it says, and the agent composes a screen for each question from a library of "hundreds of travel-specific interface elements — maps, photos, comparisons, calendars, itineraries." And because Airbnb holds the reservation and the payment, it will help during the trip with changes, payment issues and refunds, adding one sentence: "These are things an outside agent can't easily do on its own." There are support numbers too. In Q3, approximately 40% of all customer support issues were resolved without a human agent (about a third in Q2); the AI assistant is available in more than 50 languages and is expanding to voice support in the US; and customer support cost per booking fell more than 17% year over year.

We checked CEO Brian Chesky's remarks on the call against two transcripts (SingJuPost and Earnings Whispers). The outside agents' scorecard: "We're not seeing a lot of traffic and even fewer bookings from Muse and Instinct yet." "A very nonmaterial number of bookings happen on Muse and Instinct." That traffic peaked in early October and has been stable for about four weeks appears in only one transcript. He said "today's AI agents are good at simple routine tasks," "I don't think the general purpose agents are the best to do that," "We don't want to build a general agent," and "as they get really, really complicated, specialization's going to win." He does not see outside agents as enemies: "kind of like Google, an opportunity for more people to discover Airbnb," "a great handoff to us," "I don't see them as a threat." He also opened a door in the opposite direction. Asked by a Mizuho analyst whether the agent could help with travel beyond Airbnb's walls, he answered: "We're going to be integrating a number of MCPs with our trip agent in the first half next year." According to a September Skift report, Airbnb has stayed out of booking inside ChatGPT (Booking.com and Expedia joined) and sees outside assistants as "channels that already send travelers to Airbnb" (Skift, one source). We could not open the shareholder letter PDF; the sentences above are only what we confirmed on the newsroom results page and in the two transcripts.

3. Ramp — US business AI payments stuck at 55.9% for two months (facts)

We opened the Ramp AI Index from corporate-card and bill-pay company Ramp. The method: across more than 85,000 US businesses on Ramp, a firm counts as "adopting" in a month if its card, invoice or ACH payments include any positive amount to an AI product. Ramp itself writes that free tools and personal accounts are not captured, so "it's likely our results underestimate actual AI adoption and spend." Adoption in September 2026 was 55.9%, up 0.6 points from August (55.2%). Adoption by vendor: Anthropic 43.8%, OpenAI 40.6%, xAI 4.1%, DeepSeek 0.3%, Mistral 0.2%. Token-spend share (week ending October 23) was Anthropic 60.5%, OpenAI 35.1%, xAI 2.2%, Google 1.9%, but these come from a separate sample of firms using the "Token Spend Management" feature, which Ramp says "is not representative of Ramp's total AI spend volume" (see the disclosure above). In that same sample, token spending in the week of October 23 rose 8.9% and token volume 37.1% versus the same week a month earlier.

From the October 14 commentary by lead economist Ara Kharazian: "we've recorded two straight months of virtually no growth in token spend" (August and September). AI spend in early October set a new peak 1.5% above the August peak, but he still called it a significant growth slowdown. Monthly AI spend per employee at the top 1% of firms was $9,568 in September, down 3.1% from the July high of $9,873 (with the caveat that a small group makes the figure volatile), and about 80% of spend on OpenAI and Anthropic comes from 1% of firms. His reading: "It feels like we're at the top of the S-curve," meaning growth has shifted from new firms starting to firms already using AI using it more deeply; the slowdown is token spend, not seasonality; and the main driver of falling blended prices is frontier price cuts. On November 6, TechCrunch wrote of the same data that token spend rose 42% month over month in July and then only 1.5% across August and September combined, and that the median firm spends $12 to $15 per employee per month and the top 10% spend $600 to $1,000 (Ramp figures as relayed by TechCrunch; we could not find them on Ramp's pages). In Kharazian's words, "Almost the entire spend on AI is from venture-backed, high-growth tech companies." That Ramp's sample is itself tilted toward Ramp's customers — startups and tech companies — is a caveat we add, not something Ramp wrote.

Reading the three numbers together (interpretation)

First, three report cards for outside agents are in, and all three are small. Uber's four requests (NO.40), DoorDash's hundreds (NO.35), Airbnb's "very nonmaterial." In NO.38 we counted the "front doors through which AI brings customers," but the three platforms' numbers say that customers who come through that door browse but rarely pay. Chesky's word "handoff" is exact: the agent brings the customer, and the payment ends on the platform's screen (interpretation). So all three platforms react the same way. Having blocked outside agents or called them small, they build their own agent: Amazon's Alexa by text, DoorDash's text agent, Airbnb's trip agent in 2027. The door agents come through is not opening; it is changing owners.

Second, the door's conditions are being written in terms of service, not in law. The appeals court said "the one accessing is the user," and Amazon promptly blocked Muse under its terms. Amazon's two conditions — identify yourself, and respect the owner's decision about participating — are also the rules its own Buy for Me follows. In NO.37 we read one robots.txt line, and in NO.42 we separated "allowed" from "actually got in"; now there is one more column: "Did the agent that came in say who it was?" Airbnb attaching MCPs to its own agent is the other side of the same door. The direction in which the platform's agent calls outward is open; the direction in which outside agents call the platform comes with conditions (interpretation). Whichever side a small team like ours stands on, these two conditions will be the first thing we write.

Third, the number of companies using AI is nearly full; the difference now is depth. Ramp's 55.9% rising only half a point a month for two months while token spend stands still means "whether you use AI" no longer separates companies (interpretation). What separates them is where they sit between the median's $12–15 and the top 1%'s $9,568, and what they run with the money. In NO.30 we wrote "the price you pay is halved," and in NO.41 "a cheaper door changes the rate card right away"; Ramp's numbers are the result. A month has arrived in which token volume grows 37% while spend grows 9%. We are in the stretch where prices fall as fast as usage grows, so spend alone misreads as "using AI less." The caveats stand: Ramp's sample tilts toward tech companies, and Anthropic's lead is a number from that sample.

If it were us — three things for a team running apps, commerce and AI channels this week (practice)

Write three lines of "rules for incoming agents" for our apps and commerce first. Use what Amazon demanded of Muse as a mirror: ① identify itself as an agent in its requests; ② never store customers' passwords or payment methods; ③ respect it when we say we are opting out. Put these three lines, as one paragraph of terms, next to the robots.txt "three-bucket decision table" we deferred in NO.37 and the "allow list" in NO.42's practice item 2. They are not sentences for keeping agents out but conditions for letting them in. As Amazon did, we first check whether our own agents acting on someone else's behalf (the automation that drafts Ria's scripts or gathers prices) follow the same three lines. Whether this goes into our terms is the CEO's decision.
Split payments into three rows: "customer directly / brought by an agent / completed by an agent." Airbnb counted traffic and bookings separately, which is why it could say "there is traffic but bookings are very nonmaterial." We split our commerce and app payments the same way: done by the customer directly; brought by an agent (an outside assistant like ChatGPT or Muse) and completed on our screen (a handoff); and completed by an agent without our screen. The third row is probably 0 right now, and writing down that 0 is this week's job. It goes next to the "how many seconds if the customer does it themselves" column from NO.40's practice item 3. Only by keeping the method fixed for four quarters can we write a sentence like Chesky's "stable for four weeks after the peak."
Record our AI spending Ramp-style, in two columns. One column for monthly "AI payments per employee," one for "token volume" the same month. As Ramp showed, the two move differently, and watching spend alone misreads a price cut as "using less." Put it in the same table as the Gemini Omni rate re-measurement from NO.41's practice item 3. The median $12–15 and top-10% $600–1,000 are numbers from a tech-tilted sample, so we use them only for a sense of the order of magnitude, not as a yardstick for where we rank. Keeping two rows, "company announced / what we confirmed," is the same as in NO.34.

What we did not pick this time

TechCrunch reported that on November 6 Meta added "Spaces" to Muse (trackers, planners and project boards created by conversation, kept up to date by connecting outside apps) and on November 5 opened Muse Voice, letting US users video chat with their Muse avatar. We could not find an official Meta post for either, and the primary source is Mark Zuckerberg's X post, so we hold them for a later piece. The "10 million Muse users in the US and Canada" that Meta reportedly gave on its October call is also a single TechCrunch relay, so we did not use it in the body. The valuation, credit-rating and fundraising reports about Anthropic and OpenAI in the Ramp article are investment material and were not used. "Hundreds of AI hacking tools found publicly exposed" on TechCrunch's list is security news, so we recorded only the headline. The GPT-6.2 Sol routing-change claims, OpenAI "continual learning" rumors and the OpenAI developer account's "Tomorrow" teaser on promppy are sourced to X posts, so we did not register them (we will check official posts if an announcement comes). Learn:US's two November 5 posts (replica skills, Firecrawl connector) we saw only in the list summary and hold as follow-up candidates to NO.26 and NO.32 on "checking before installing."

What we could not confirm in this news

Amazon's "400 million" and "eightfold" are undefined (whether a single use counts; the scope of "agentic features"). The September 20 block date and Amazon's reasons agree across two outlets, but Amazon's statement, Buy for Me's self-identification and opt-out, and the restrictions on Google and OpenAI agents are relayed via GeekWire. We did not open the identifying-string clause in Amazon's terms. DoorDash's "focus on its own text agent" is TechCrunch alone.
We could not open Airbnb's shareholder letter PDF; we confirmed via the newsroom results page and two transcripts. "Traffic peaked in early October and has been stable for about four weeks" is in one transcript only. Which company makes Instinct is not in the transcripts, so we described it only as "a personal agent app Chesky named alongside Muse." Staying out of ChatGPT booking and Booking.com/Expedia joining are Skift alone.
Ramp's median $12–15, top-10% $600–1,000 and July +42% are figures relayed by TechCrunch that we could not find directly on Ramp's pages. The token-spend share (Anthropic 60.5%) comes from a separate sample Ramp itself calls "not representative." That Ramp's sample tilts toward tech companies is a caveat we drew from Kharazian's remark; Ramp has not published its sample composition.
The Perplexity case timeline (filing, injunction, appeal, rehearing denial) is PPC Land's summary; SiliconANGLE gave only the appeal's conclusion. We did not open the court opinions.

What we read alongside

We opened sources of different kinds directly. We list company announcements, payments data, earnings call transcripts, media coverage and our previous articles separately.

Company announcements:

Payments data:

Earnings call transcripts:

Media coverage:

Our previous articles: