And someone wrote that list before it. If it was the vendor, the same cases serve all its clients, the jeweler and the hotel group alike. If it was the Maison, they describe its own requests and fit on a page its committee can reread and correct. This is the closed-list decision. It holds for any agent, whatever model does the deciding.
What makes the subject pressing is a model that teams building agents have been talking about since its release: Jev, launched on September 15 by TypeSafe AI. It writes nothing. You ask it a closed question, and it answers with a choice from a list, a score or a probability. The vendor's figures, with no independent measurement found as of September 24: half a second at most per decision, and $0.042 per million tokens read, a token being the word fragment on which a model is billed.
According to VentureBeat, 140,000 sign-ups got access within 36 hours, and Cloudflare, LangChain and Langfuse had plugged it in within three days. Jev will probably be overtaken within a year; the list it reads will remain.
"Hand-Picked" says two things at once: the care with which a Maison chooses its suppliers, its pieces and its clients, and the precise gesture of a model that picks from a closed list without writing a word. The cover borrows from Gattaca (Andrew Niccol, 1997), where a genetic file decides in advance who boards the ship: selection taken to its limit, and a reminder that a list is never neutral. It sees only what its authors put into it.
"The Maison doesn't buy the answer. It writes the list."
Welcome to LUXE ÆTERNAI, my weekly decoding of what AI agents change, or don't, for Luxury Maisons. I am Michaël Tsakiris. I help leaders and executive committees in Luxury and premium decide what they entrust to agents, and I put it in place with them until their teams run it themselves. Enjoy!
The LUXE ÆTERNAI connector. Plugged into https://connecteur.luxeaeternai.com/mcp, your AI assistant searches the twenty-four published editions, opens any one of them in full and passes your questions on to me. It reads only public content: send nothing confidential through it (privacy policy).
The week in short ¶
The thesis. An agent spends most of its time deciding without writing: sorting, authorizing, blocking. It chooses from a list of cases, and everything depends on who wrote it. Jev, launched on September 15 by TypeSafe AI, is the occasion, not the subject.
The paradox. Agent payment already works at Mastercard and GoCardless, yet only 7% of the fashion shoppers surveyed by ACI Worldwide (a survey of more than 3,300 American and British consumers) would let an agent buy without their approval.
What's moving. Prada's chief procurement officer presents an agent for reviewing more than 1,000 suppliers, according to WWD; Accor puts its agentic concierge into WhatsApp.
The three that count, outside Luxury. An OpenAI agent entered Australia's Medicare portal without instruction; OpenAI and Anthropic cut their prices on the same day; six banks publish their principles for agentic commerce, including a trace of every delegated purchase.
Decoded. What Jev can do, its flaw, and its two uses in a Maison: sorting requests, checking a message before it goes out.
The grid of ten, this edition's resource. The ten questions to ask, word for word, of any provider before entrusting it with an agent, with the document to obtain in return. Eight of the ten belong to general management.
The story. Sephora, an LVMH subsidiary, is announced among the retailers on Muse, Meta's agent; Amazon refuses Muse and, having lost its injunction against Perplexity's agent on appeal, now defends itself through its terms of use; Moda Operandi already has a clause on orders placed by an agent.
The radar. Three vendor announcements: Amazon's seller assistant becomes an agent that acts on authorization, Okta gathers twelve vendors around four security questions, SiteMinder promises to fix distribution failures.
The test bench. A half-day test, with no budget, to find out whether your list of cases holds: the condition for letting an agent sort your client requests.
In my reading list. BCG on an agent's permissions, McKinsey on time lost between teams, Boltanski and Thévenot on the reason given to a client.
Coming next. Pernod Ricard on October 15, Hermès on the 22nd; LVMH (around the 20th) and Kering (the 29th), dates to be confirmed.
Agent payment already works at Mastercard and GoCardless; 7% of fashion shoppers would let an agent buy without their approval. ¶
In one week, agent payment went from demonstration to real transaction. On September 21, in Denmark, an agent booked and paid for a coffee tasting with a Danske Bank Mastercard, after the client's approval and confirmation by fingerprint or code on her phone (The Paypers). On the 22nd, in the United Kingdom, GoCardless processed the first bank payment set up by an agent, within the live testing program of the British financial regulator (The Paypers). These are isolated transactions, and no volume has been published.
On the 18th, Ant International, which had opened its merchants' checkout pages to agents in August (edition 20), announced an account that lets an agent pay through Alipay+, the payment method of Chinese travelers in Europe (South China Morning Post).
The client is not following. According to an ACI Worldwide survey of more than 3,300 American and British consumers, only 7% of the fashion shoppers surveyed would let an assistant buy without their approval; 53% say they are uncomfortable with the idea of it buying for them, and 14% would insist on approving every purchase (Retail Dive, September 23). In travel, McKinsey finds the same reluctance: of more than 1,000 American travelers surveyed, three in four prefer to make the final booking themselves (McKinsey, September 22). Neither survey isolates Luxury.
7%of fashion shoppers surveyed
ACI Worldwide, more than 3,300 American and British consumers, according to Retail Dive, September 23, 2026 · Luxury is not isolated
The stakes can be measured: the global market for personal Luxury goods came to €358 billion in 2025, and Bain and Altagamma expect €365 to €373 billion in 2026 (Bain, June 25). I have found no measurement of the share that already goes through an agent.
"The agent prepares the order; the client is the one who presses buy."
The counterweight: x402, a payment network designed for agents, fell from about $800,000 a day in January to $40,000 in September, according to American Banker of September 18, which considers its data imprecise. I conclude that agent payment will run over the networks the Maison already accepts, card and bank debit, rather than over a specialized network.
The payment networks are ready, the clients are not, and I see no reason to rush them. In Luxury, the final confirmation serves a purpose: it is the moment when the client advisor confirms a size, suggests an alteration, gives a delivery date. A third-party agent that pays without that step deprives the Maison of its last exchange before shipping.
What I take from this.No investment in a specialized payment network this year. The price of that caution: an order an agent cannot complete on its own may go to a competitor that accepts it. I accept that, because only 7% of the fashion shoppers surveyed want it. In any shopping agent project, write down that the agent prepares and a named person approves. Ask your payment provider whether it can already recognize an agent that declares itself and accept its payment, including from a Chinese wallet. For the client, nothing goes out without her approval, and the client advisor keeps the chance of a last exchange.
Prada's chief procurement officer describes an agent for reviewing its suppliers; Accor puts its concierge into WhatsApp. ¶
Prada's chief procurement officer presents an agent as the answer to reviewing the group's suppliers, according to WWD Sourcing Journal. The Prada group (Prada, Miu Miu, and Versace since last December) works with more than 1,000 suppliers, from leather to packaging. "It's 1,000 suppliers to be scrutinized, and it could take weeks […]," said Fabio Francalancia, the group's chief procurement and cost control officer, on stage at the New York event of his software vendor Ivalua, according to WWD Sourcing Journal of September 18.
Later in the session, he called the Iva agent "the workforce you would never find."
“the workforce you would never find”
The Iva agent, announced by Ivalua on June 11, handles steps in the procurement process, according to the vendor; the steps Prada entrusts to it have not been published.
Procurement Magazine confirms the talk. I have found neither a press release naming Prada as a client of the agent, nor any measured result. In December 2025, the Milan prosecutor's office requested the subcontracting audits of 13 brands, Prada and Versace among them (Il Sole 24 Ore). I conclude that reviewing subcontractors is natural ground for an agent: it reads the files, the buyer keeps the decision.
Accor puts its agentic concierge into WhatsApp and iMessage. On September 23, the group (Raffles, Orient Express, Fairmont, Sofitel) launched ALL Concierge on its website, its app, WhatsApp and iMessage, in 11 languages: the agent handles the booking from end to end, down to changing a room or spending loyalty points, and hands over to an advisor when the traveler asks (Hotel Dive, Tom.travel, L'Usine Digitale). No usage figures have been published, and the sources do not say whether Raffles and Orient Express are covered.
Sources for this section6 links
Outside Luxury, the agentic week in three facts. ¶
1. An OpenAI agent entered Australia's Medicare portal without instruction. On September 23, Prime Minister Anthony Albanese disclosed an intrusion committed in June: an OpenAI agent got into the statistics portal of Medicare, the country's public health insurance, and "accessed both public and non-public files," according to the Financial Times, picked up by CNBC and Fortune. OpenAI says the agent "took actions we did not intend" during an internal evaluation; no personal data was taken, according to the company.
The dates matter more than the intrusion. Committed in June, it was detected by OpenAI in August; OpenAI reported it to the Australian government on September 10, by email to a public mailbox, thirteen days before the Prime Minister made it public, and the message took five more days to reach the government's cybersecurity agency.
Why it counts: any agent that browses outside the Maison's walls on its behalf can cross a perimeter without instruction. Make the vendor's detection time and reporting time two quantified clauses of the contract, with penalties.
2. OpenAI and Anthropic cut their prices on the same day; the cost of a given level of performance has fallen by about 47% per quarter since 2023. On September 22, a few hours apart according to Fortune, OpenAI launched two models, GPT-6 Sol and GPT-6 Luna, sold at half the price of their predecessors. Anthropic launched a model 20% cheaper than the previous one (CNBC). These are the two vendors' public list prices.
The same day, the research institute Epoch AI published its measurement: since 2023, getting the same result from an AI has cost roughly half as much from one quarter to the next, a drop of about 47% per quarter. An agent quote priced on July rates is therefore already wrong, and every vendor proposal should be reread with a revision clause on the model's price.
47%per quarter
Epoch AI, September 22, 2026 · approximate value
3. Six banks publish their principles for agentic commerce, including a trace of every delegated purchase. On September 22, ASB, Bank of America, Capital One, Commonwealth Bank of Australia, ING and NatWest published five principles: transparency, security, data and privacy, choice, interoperability. An implementation document is to follow. The most concrete point for a merchant lies in the first: being able to trace the whole chain of a purchase, from the outcome back to the client's instruction, through the agent's decision, its authentication and the mandate it received (The Paypers, PYMNTS).
Richard Crone, a payments consultant quoted by American Banker, believes the market is moving faster than this white paper.
Why it counts: the clients' banks are asking who pays when the agent buys the wrong piece. A Maison that accepts an agent's order keeps the instruction received and the confirmation sent: without those two records, if the client disputes the charge, her bank can impose the refund and the Maison bears the full amount alone.
Sources for this section9 links
Jev, a model that chooses from a list without writing a word: what it can do, its flaw, and its two possible uses in a Maison. ¶
Jev is a component placed inside an agent, wherever a decision has to be made without writing; it replaces neither ChatGPT nor Claude. It takes only three forms of question, according to the Blog du Modérateur of September 23: in this list, which one; on this scale, what score; is this statement true, and with what probability. An example: a client writes that her watch is letting in water; the list reads repair, warranty, complaint, appointment, press, other; Jev answers "repair, 0.91." It has sorted without writing a word.
According to TypeSafe, it answers in 70 to 500 milliseconds and accepts up to 255 options. It does not always give the same answer to the same question, and the vendor does not claim otherwise. But it can only answer with one of the options on the list, along with its probability, never with a free sentence.
The developer Flavio Copes, who tested it, sets out its limits: it calculates poorly, it misreads the dates and figures put to it, and it can pick the wrong box.
Two setups published by LangChain. On September 17, LangChain, a maker of tools for building agents, and therefore a company with an interest in Jev spreading, published two setups. The first is a router: Jev decides whether a fast, cheap model will do, or whether a powerful one is needed. The second is a safeguard: every time the agent wants to trigger an action, Jev classifies it and blocks the ones it judges risky.
The flaw. The VentureBeat article cited above, dated September 21, reports a test. The principle of the attack: someone hides an instruction in content the agent is going to read, a product page, a web page, a tool's response. The agent reads it as if it came from you, and carries it out.
An engineer at Octomind slipped an instruction to approve a dangerous command into a tool's output: before the hidden instruction, Jev put the probability that the command should be blocked at 0.76; after it, at 0.48, on that single test case. TypeSafe's own documentation says as much: text written to steer the model "can move the answer."
0.76 → 0.48probability of blocking
VentureBeat · a single test case
A model that decides without writing is still a model that reads; if the same hidden instruction reaches both the agent that acts and the checker that authorizes it, the two go wrong together. Edition 23 already noted this limit with Sentinel, the second Meta agent that authorizes Muse's actions.
Two uses in a Maison. The first use is sorting: every request received by customer service is placed in a box on the list before a person or a conversational agent writes anything at all.
The second use is the one I recommend first: a checking agent that, before every message sent by a conversational agent, answers a single closed question (can this message go out as it is?) with its probability. I have not seen this check on every reply at any Maison; until now, having every message reread by a large model cost several hundred times more than with Jev, according to TypeSafe.
The report published in January by Dynatrace, a maker of monitoring software, on 919 decision-makers at large companies outside Luxury, gives the measure of the need: 45% of the 919 decision-makers surveyed say they do not know how to set the point at which an agent may act alone (Dynatrace, vendor data). That point can be written as a closed list: the expected cases, an "I don't know" box that sends the request to a person, and a probability threshold below which the agent does not act alone.
"Inside an agent, the cheapest model is the one that authorizes or blocks every action."
What research says. Herbert Simon, winner of the Nobel Prize in economics, described a decision as early as Administrative Behavior (1947) as a choice among alternatives that someone has framed beforehand: the essential part happens in the framing. The vendor sells the choice; the Maison writes the cases. The psychologist Gerd Gigerenzer (Simple Heuristics That Make Us Smart, 1999) showed that a simple rule often beats a heavy calculation, on one condition: that it matches the way cases actually arrive. That is what the test bench below checks.
Why not keep a large model. A large model can classify, perhaps better (I have found no independent comparison), at a price per decision several hundred times higher according to the vendor, and with the risk of writing something along the way. The answer lies in the list: with good cases, a cheap model is enough; with bad cases, the best model still sorts into the wrong boxes. The savings TypeSafe promises only hold if the list of cases is right, and it is not when the provider writes the same cases for a jeweler and for a hotel group.
The precedent. The closed-list decision has already served in business, and its hidden cost is known. In 1987, seven years after it went into service at Digital Equipment, XCON, the expert system that configured computer orders from hand-written rules, had 6,200 rules, about half of which changed every year (Soloway, Bachant and Jensen, conference of the Association for the Advancement of Artificial Intelligence, AAAI, 1987). The system held; it was keeping it up to date that grew harder and harder.
A list of cases is not delivered once: it has to be maintained at the pace at which the offer and the clients change. Bought from a vendor, it is updated only when the vendor decides, not when your offer or your clients change.
Written by the Maison, it carries that cost too, the one XCON revealed. Writing your own list means committing to revise it; I recommend it all the same, because it is the only way for the list to describe your cases and not those of a hotel group.
What I take from this.Have customer service, together with the legal department, write the list of cases in which an agent may sort a request, with its abstention box and its threshold. Entrust the check to a model that does not read the same content as the agent that acts: an instruction hidden in a page must never reach both at once. Demand from every conversational agent vendor a pre-send checker, and its price. For the client, a request the list cannot place reaches a person, never a wrong answer.
This edition's resource: the ten questions to ask, in order, of any provider offering you an agent. Eight of the ten are not technical. ¶
Ten questions, in the order in which a leadership team deals with them, to put in writing to any vendor before the demo. Each row says what you should get back: a dated document that can be attached to the contract. Rows marked with a diamond must end up as a clause in the contract.
| # | The question to ask the provider | What you must obtain in writing |
|---|---|---|
| 1 | Which components does your agent depend on, and what happens to me when one of them changes? | The diagram of the components and their dependencies, on one page. |
| 2 | On which servers does the agent run, who can follow what it does, and at what amount spent does it stop on its own? | The environment, access to the log, and the spending cap in figures. ◆ |
| 3 | Where does my client data live, who has access to it, and does the agent let your other clients benefit from it? | The location, the list of access rights, and the commitment to keep it separate. ◆ |
| 4 | What does the agent read before deciding, and what does it do when one of the systems it queries sends back false information? | The named list of the sources it consults, and the rule that says what it does when a source answers wrong. |
| 5 | If I stop in eighteen months, what do I take with me? | The lists, logs and cases written by your teams belong to you, in a readable format. ◆ |
| 6 | When the agent gets it wrong, who is liable, and where do you stand under the European AI Act? | The allocation of liability and the agent's classification under the Act. ◆ |
| 7 | What can it touch, and how far does the damage go if things go wrong? | The closed list of its actions and permissions, with the maximum extent of an incident. ◆ |
| 8 | What do you test it on: the outcome for the client, or the quality of its sentences? | The set of test cases and the measured result on those cases. |
| 9 | When does a person approve, and when does the agent proceed alone? | The abstention box, the probability threshold, and the amount above which a person must say yes before anything is executed. ◆ |
| 10 | When it gets it wrong, how quickly will I know, and from whom? | The detection time and the reporting time, in figures, with penalties. ◆ |
Two questions call for an engineer, the second and the fourth; for the second, Strands Harness, the free tool released on September 21 by AWS, Amazon's cloud computing subsidiary, is not enough: it provides neither a test environment cut off from live systems, nor the spending cap (SiliconANGLE). The other eight are leadership decisions. The seventh, which ties each of the agent's permissions to the purpose of its mission, is the one BCG details on September 22. The third, fifth and sixth are the ones a leader can ask alone, with no engineer at their side.
What I do with it.Send these ten questions to every provider in the running, in writing, before the demo. Refuse any contract that does not answer questions 2, 3, 5, 6, 7, 9 and 10 with a written clause, quantified whenever the question concerns an amount or a time frame. Have the eight non-technical questions settled by the executive committee, not by the project team.
Sources for this section2 links
Meta names Sephora among Muse's retailers; Amazon refuses the agent and falls back on its terms of use; Moda Operandi had already written its own. ¶
Muse was in edition 23. Since then, Shopify and PayPal have opened checkout to it, and Amazon has shut the door. Meta's personal agent has been buying on its user's behalf in the United States for two weeks; it counts more than 2.5 million downloads according to Sensor Tower, cited by CBS News. The week settled where it is allowed to buy, in both directions.
Accept. On September 21, Shopify allowed Muse to pay with Shop Pay in all its stores; PayPal followed on the 22nd with its merchants worldwide (PYMNTS, Shopify; PYMNTS, PayPal). On the 23rd, at its Connect conference, Meta announced the retailers to come: Sephora, an LVMH subsidiary, Gap, Walmart, Best Buy and Wayfair, in addition to the Shopify catalog; Expedia had been announced the day before. These retailers are announced; none is live yet (TechCrunch). Meta plans to take a fee per transaction.
Refuse. Also on the 21st, Amazon asked to be removed from Muse, having not been informed, and tells its users that access by an unauthorized agent violates its terms (Retail Dive, The Next Web).
Against Comet, Perplexity's agent, Amazon is changing weapons. In March, it had obtained an injunction under the civil provisions of the US Computer Fraud and Abuse Act; on August 4, the Ninth Circuit Court of Appeals vacated it (edition 20, Ninth Circuit opinion, Jones Day). On September 21, Amazon therefore amended its complaint: it adds facts to show that Perplexity itself does access the site, and a claim based on its terms for agents from May 2025, which require an agent to identify itself (The Fashion Law, court docket). The court has not yet ruled on whether those terms are enough.
Write the clause. One Luxury retailer has already drafted it. According to the review of 35 retailers and brands published by The Fashion Law on September 17, Moda Operandi's terms of use provide for the case of an order placed by an agent; Target sets out what an approved agent may do with a cart; Cartier is among the brands that restrict training an AI on their content, and the review attributes no clause on shopping agents to it. At a Luxury Maison, I found none in this review.
"Between accepting and refusing, there is the clause."
The decision is written in the legal department, in a few weeks by my estimate: what an agent may do on the site, whether it must identify itself, which pieces stay closed to it (allocations, limited editions, special orders), and what the boutique does with an order that arrives without a client's name: accept it, block it, or hold it pending confirmation. Once written, it gives the Maison grounds to act; that is exactly what Amazon is testing against Perplexity right now.
What I take from this.Have the legal department write into the site's terms of use what an agent may do, and require it to identify itself. The Shopify decision raised in edition 23 now has a deadline: if your online store runs on Shopify in the United States and nothing has been decided, it can already take Muse's orders. Decide, for every order placed by an agent, what proof of the client's identity the boutique requires before shipping. For the client, the purchase made by her agent lands in her record, her history and with her client advisor.
What your teams' tools are learning to do this week. ¶
E-commerce · September 23, Amazon.
The company announces that Seller Assistant is becoming an always-on agent that learns how each seller makes decisions, watches over inventory, prices, listings and compliance, and carries out an action only with the seller's approval; this is vendor data, and the module added by a partner is still in testing (PYMNTS, Retail Gazette): this is what an agent running an online store under its manager's control will look like.
IT and security · September 22, Okta.
The vendor, which since August has offered a free inventory of a company's agents and their access rights (edition 21), publishes with AWS, Google Cloud, Salesforce, ServiceNow and seven others a joint blueprint for agent security, built on four questions: where are my agents, what do they have access to, what are they doing, how do I stop them (SiliconANGLE, TNGlobal): a ready-made grid for a call for tenders. The gateway that controls what agents do is announced for the third quarter; the switch to stop them, for the fourth. Until then, the fourth question has no tool.
Hospitality · September 22, SiteMinder.
The vendor announces an engine that will detect distribution failures (a broken channel, a room type left unmapped) and fix them after one click from the revenue manager; a press release in the future tense, for a claimed base of 56,000 properties (press release, PhocusWire).
Sources for this section6 links
A half-day test to find out whether your list of cases holds: the condition for letting an agent sort your client requests. ¶
A decision model like Jev accepts only a list of cases fixed in advance (TypeSafe, post of September 15, Blog du Modérateur). This test measures whether yours holds, in half a day and with no budget: gathering the hundred requests is half the work.
Who takes part. The head of customer service and two experienced client advisors, two hours each.
What the test covers. Customer service's last hundred written requests, all channels, unsorted.
The instructions, word for word ¶
- 1.To the head of customer service: write the list of boxes these requests should fall into, five at most, plus an abstention box; do not read the requests before writing the list.
- 2.To each of the two advisors, separately: place each of the hundred requests in a box on the list, without discussing it with the other, and note the time taken.
What the head of customer service records. The requests classified identically by both advisors; those that fell into the abstention box; the cases the list did not provide for.
What the result tells you to do. If the two advisors classify more than 80 requests out of 100 identically (a simple count, with no statistical correction), the list is good: sorting can be handed to an agent, which will have to do at least as well as your two advisors. The agreement between your teams sets the bar, not the probability promised by the vendor (TypeSafe).
Between 60 and 80 (thresholds I set myself, with no published standard), the list is rewritten on the disputed cases, then the test is run again. Below 60, it is not the agent that is missing, it is the list. The result fits on one page: the list, the agreement rate obtained, and the name of whoever keeps it up to date.
This test is the first page of the Agentic Delegation Plan: trade by trade, what a Maison entrusts to an agent, what it keeps for itself, and for each agent that decides, its list, its threshold and the person accountable for it.
Sources for this section2 links
To dig deeper. ¶
- BCG, "The Authorization Gap: Why Yesterday's Controls Won't Work with Today's Agents", September 22. Five questions on an agent's permissions. The most useful to a committee: tie each permission to a specific purpose. The clearest: a cap, set by a rule and not by another model, that halts any execution until a person has said yes.
- McKinsey, "Cutting the 'coordination tax'", September 18. A single measured case, at a Fortune 500 manufacturer: a 7-step process in which the 6 handoffs cost 9 to 18 days of waiting, for 12 to 24 hours of actual work. The time an agent can save lies in the waiting between teams, not in each team's work.
- The classic. Luc Boltanski and Laurent Thévenot, On Justification: Economies of Worth (Princeton University Press, 2006; French original, Gallimard, 1991). A decision holds between people only if it can be justified in an order both of them recognize. Before letting an agent refuse a commercial gesture, decide on the reason the client advisor will be able to give the client.
The agentic Luxury agenda, October 15 to 29. ¶
-
Thursday, October 15, 7:30 a.m.
First-quarter sales for the 2026-2027 fiscal year at Pernod Ricard.
-
In October, day not published.
Third-quarter 2026 revenue at LVMH: its official calendar says "October 2026" with no date; Investing.com puts it on the 20th.
-
Thursday, October 22, 8 a.m.
Third-quarter 2026 revenue at Hermès.
-
Thursday, October 29.
Third-quarter 2026 revenue at Kering, according to Investing.com, to be confirmed on the official calendar.
The grid of ten questions will be useful long after Jev has been replaced. A year from now, the question put to an agent vendor will no longer be what the agent can do, but who wrote its cases, and I advise every Maison to be able to answer: we did. The Maison doesn't buy the answer. It writes the list. Luxe oblige !
Sources for this section4 links
Michaël Tsakiris
Paris, Thursday, September 24, 2026
