Paroles, paroles

• The thesis. How much a group talks about AI measures the pressure of its market, not its progress. The corollary for the season: never measure a competitor by its talk; read its acts (its job listings, its observed deployments) and take your own measurements.
• Fast rewind. The whole summer in a single timeline, from June 17 to August 18: thirteen dates, five of which shape the season.
• The story. Three earnings calls in three days: Kering narrates, LVMH answers, Hermès says nothing. Three ways of handling the same question, and not one of them tells you who is ahead.
• The pro tools radar. A new section: Google, Shopify, SAP, Akeneo, Meta and Microsoft changed, over the summer, the software your teams open every morning.
• On the test bench. The new closing section: a test you can run in-house that turns the week's thesis into a measurement for your Maison. This week: what the agents make of yours, in four runs and one page for the executive committee.
• What changes in this newsletter. The format you are reading becomes the Thursday appointment; every two weeks the LUXE ÆTERNAI Deep Dive is added to it, a long read that takes on a single subject; and a permanent section on tools makes its entrance.
• This newsletter installs itself in your favorite AI assistant. Thirty seconds, one address to paste, the how-to for each assistant: your assistant consults the corpus instead of answering from memory, and carries your questions to me.
• In the first Deep Dive (September 3). The jobs the agentic era is creating in Luxury, and the talent war that has already begun: the roles that did not exist two years ago, who is hiring what, where, at what price. The full compilation of the sector's real hiring, which no one has published.

Backstage first. This newsletter is prepared, in part, by teams of bespoke agents, among them LAE Studio, my workshop: sourcing, fact-checking, quality control, each with its role, as in a newsroom. I have worked this way for a long time; the summer put these tools in everyone's hands (GPT-5.6 on July 9, Claude Opus 5 on the 24th, prices down as much as 80%). What will not be delegated: the writing, the judgment, the analysis, the references. The agents carry the file; they sign nothing.

What it changes for you:

• The summer at speed, just below: a catch-up section, published this once only.
• An appointment every Thursday, in the format of this issue: the week's news decoded, the story, the tools radar, the test bench. And every two weeks, in the same send, the LUXE ÆTERNAI Deep Dive: a long read that takes on a single subject. The dates are at the end of the newsletter.
• One more section, permanent: THE PRO TOOLS RADAR. The notable updates to the software actually deployed inside Maisons (office suites, CRM, media buying, commerce, ERP, project management, search), read from a single angle: what it changes for your teams. And when a vendor has shipped nothing notable, the section writes that, instead of padding.
• A test to run in-house in every issue: THE TEST BENCH, at the end of the newsletter, with the full procedure and the findings to take up to your executive committee.
• This newsletter installs itself in your favorite AI assistant, and becomes your channel for answering me. It is the change that needs the most explaining: the next section is devoted to it.

Thirty seconds, no technical skill required, and your assistant stops answering from memory on these subjects.

What this is. Asked about agentic Luxury, your AI assistant improvises with what it retained from its training: approximations, stale figures, sometimes an invention. The connector gives it access to my archives: it goes and reads them, then answers you citing the issue and the date. You get a traceable answer instead of a plausible one.

What it consults, in full:

• The twenty issues published since February 2026, plus the special edition, in full text; the corpus grows with each new issue, on Thursdays: luxeaeternai.com/en/newsletter
• The study "The client, the brand, and the agent" (April 2026): 680 answers from four AI assistants on ten Maisons, 83 pages: luxeaeternai.com/en/study/the-client-the-brand-and-the-agent
• The AI × Luxury Glossary, updated issue after issue: luxeaeternai.com/en/glossary
• The site luxeaeternai.com
• My working reference base, the one I keep in Notion to prepare every issue and that is published nowhere: 139 Maisons and players tracked with what they actually do in AI, 63 concepts, 211 sources and studies, 66 terms.

You change nothing in your habits: ask your questions as before; it, from now on, checks at the source.

The address to paste, the same everywhere, exactly as it is and without quotation marks:
https://connecteur.luxeaeternai.com/mcp

How to install it, depending on your assistant:

• Claude (all plans): Settings, then "Connectors," then "Add custom connector." Paste https://connecteur.luxeaeternai.com/mcp and confirm.
• Perplexity (Pro, Max or Enterprise): Account settings, then "Connectors," then "+ Custom connector." Choose the "Remote" type, paste https://connecteur.luxeaeternai.com/mcp and confirm.
• Mistral's Le Chat: "Connectors," then "Add Connector," then "Custom MCP Connector." Paste https://connecteur.luxeaeternai.com/mcp and confirm.
• ChatGPT: custom connectors are open only to Business, Enterprise and Edu plans, and it is your IT administrator who turns them on. Pass them https://connecteur.luxeaeternai.com/mcp; in the meantime, the whole corpus stays readable on the site.

What you will be able to ask it next, in plain language: "what did LUXE ÆTERNAI say about sales copilots?", "summarize the issue on brand sovereignty for me", "what is the agentic footprint, according to Mickaël Tsakiris?". It will go and read, then answer citing the issue and its date.

And to write to me: your agent talks to mine. The connector runs both ways. Tell your assistant "send this question to Mickaël": it checks with you what it is sending, then delivers it straight to my system, where my own agents sort it before I do. Agent to agent, literally; the letters page has just changed century. You stay in your conversation, with no form and no address to hunt down, and I answer you in person. Prefer your thumb to a keyboard? The newsletter has its WhatsApp too: save +33 7 45 58 18 71 in your contacts, or tap wa.me/33745581871 from your phone, and the conversation opens. Write as you would to a contact: I read everything. The most illuminating questions feed the coming issues, always anonymized.

The whole summer in a single timeline, from June 17 to August 18. The five facts that shape the season are developed; the rest fits on one line.

June 17. L'Oréal puts Maybelline try-on inside ChatGPT; beauty takes a lead on Luxury inside the assistant.

June 30 · The founding figure of the summer. 82% of the highest-spending Luxury buyers used an AI during their last purchase; across from them, 22% of Maisons rank AI among their top three priorities (Bain & Company × Comité Colbert study). For the first time since e-commerce, it is the client, and the most valuable client at that, who sets the pace of adoption of a technology. Everything else this summer reads through that gap.

July 9. The UK competition regulator equips itself with its own agentic AI to spot consumer harm at scale. The watchdog is tooling up before the brands.

July 15 · China draws the line on the acceptable agent. ByteDance and Alibaba unplug their agents that simulate an emotional relationship, regulation obliging. In Luxury's leading adoption market, a virtual advisor will have to be an agent of service, not a simulated relationship.

July 21. Microsoft and Mistral extend their partnership all the way to fully disconnected deployments; the sovereignty argument reaches the European Maisons.

July 28. Radisson puts more than 1,000 hotels inside ChatGPT, live inventory and rates.

July 29. More than a million businesses a week use Meta's Business Agent, which answers, recommends and sells on WhatsApp and Instagram; paid from August 1. The king channel of clienteling goes agentic.

July 30. Amazon puts numbers on its shopping agent, Alexa for Shopping: 350 million users in twelve months, and a basket per order 40% higher among its US users.

August 2, and the reversal almost no one flagged. The European AI Act was due to come into full application for high risk. Six days before the deadline, an omnibus regulation (EU 2026/1744, published July 24, in force July 27) pushed those obligations to December 2, 2027, and to 2028 for AI embedded in regulated products. The scale of penalties cited in #19 is therefore not enforceable for high risk. Two obligations did take effect on August 2: telling the customer they are talking to a machine and marking generated content (Article 50), and training your teams in AI (Article 4). Brussels postponed technical compliance and kept the competence requirement.

August 4 · US courts make the agent an extension of the client. The Ninth Circuit Court of Appeals vacates Amazon's injunction against Perplexity's shopping agent: it is the human user, not the software vendor, who is deemed to be accessing the site. The case continues, but blocking third-party agents has just lost its most convenient legal footing. The client's agent will come in; the terms of welcome are yours to write.

August 4, again. Zalando's assistant now searches the web and recognizes a garment in a photo; the interactions the group considers most profitable rose 63% over the half.

August 17 and 18 · Credit and loyalty migrate into the assistant. Synchrony plugs financing and loyalty into ChatGPT's checkout; the next day, more than 25 players including Visa and Mastercard join forces to standardize agent payments. After payment, loyalty, the sinew of clienteling, settles where the agent buys.

August 18 · Alipay industrializes agent commerce. Alipay hands merchants what they need to turn their web pages into functions agents call directly, without going through the site. Tomorrow's Chinese "site" is a set of tools; whoever does not publish in that format does not exist in the journey.

Sources:

• L'Oréal — L'Oréal × OpenAI, Maybelline virtual try-on in ChatGPT, June 17 ; Bain & Company × Comité Colbert — "Winning Over the Customer in the Age of AI", full report, June 30 ; CMA — Annual Report 2025-2026, agentic tools and Internet Lab, July 9 ; South China Morning Post — ByteDance and Alibaba disable humanlike agents, July 15 ; Microsoft — expanded partnership with Mistral, July 21 ; Accenture — Radisson Hotel Group on ChatGPT, July 28 ; Unite.AI — Meta Business Agent, more than a million businesses a week, July 29 ; Retail Dive — Amazon, Alexa for Shopping, July 30 ; European Commission — AI Act ; Regulation (EU) 2026/1744, Digital Omnibus, published July 24, in force July 27 ; Ninth Circuit Court of Appeals — Amazon v. Perplexity, opinion of August 4 ; Zalando — first-half results, August 4 ; Synchrony — collaboration with OpenAI, August 17 ; PR Newswire — Agentic Payments Alliance, August 18 ; Crowdfund Insider — Alipay, agentic commerce infrastructure, August 18.

From July 27 to 29, in three days, the three largest French Luxury groups presented their half-year results. Three calls, the same analysts on the line, the same technology on the table. Listen to the tapes in order: these are not three AI strategies in opposition, they are three ways of keeping quiet.

On July 27, LVMH speaks of AI only when invited to. The group's results release does not contain the words "artificial intelligence" a single time. During the call, management does not say them either; it takes an analyst, Oliver Chen, to put the AI question on the table. The chief financial officer, Cécile Cabanis, then answers calmly: the group uses it for clienteling ("how you can really make the relationship and the conversion much more efficient," she says), to forecast demand and to speed up prototypes. Then she moves to the next item. Savor the paradox: the world number one in Luxury, €38.6 billion in revenue for the half, uses AI every day, and feels no need to say so as long as no one asks.

On July 28, Kering puts AI at the heart of its earnings call. Luca de Meo presents analysts with an AI program named "Augmented Kering": agents the group is preparing to deploy not facing clients, but inside its own processes, to speed up the execution of its plan. He sums up the logic in one sentence: "We follow the money, we follow the product, and we follow the client." Three priorities: costs, collections, clientele. That same day, the group announces its first quarter of growth in three years; the stock gains nearly 16% on the day. Everything is put forward that morning: the program, the three axes, the return to growth. Everything, except the detail that proves the rest is serious: the hiring had started four days before the call.

On July 29, Hermès does not say the word AI once. Not a line in its results, not an analyst question that presses, not a word to quote back. Complete silence, at the Maison whose valuation could not care less about the technology tempo and whose discourse has not shifted an inch in ten years. Chanel, unlisted, does not even have to walk onto that stage: it publishes its accounts when it has decided to, and has never commented on a software roadmap in front of anyone.

The lazy ranking, and why it is wrong. The reflex would be to conclude: Kering is ahead, LVMH is buying time, Hermès is asleep. Wrong three times over, in all likelihood. An earnings call exists to convince a market. It does not exist to inventory deployments. A group in turnaround needs a story of transformation; a group with nothing to prove that quarter spares itself a promise, because a promise binds and silence costs nothing. De Meo announces agents he will deploy; Cabanis describes uses already in place, without putting them in the window; Hermès says nothing because nothing obliges it to. How much a group talks about AI does not measure its progress: it measures the pressure of its market.

An earnings call is a stage. The sociologist Erving Goffman called it face: what one means to preserve in a public interaction. An earnings call is a ritual of face, not an audit. What gets said there about AI belongs to the same work of self-presentation as the choice of a magazine cover, and no one, on a trading floor, has ever asked to see proof of a deployment.

What the groups keep quiet at the microphone, their job listings say out loud. Career sites are public, dated and free: I made the count there on August 24, group by group. Three days after its call, on July 31, Kering opened a "Data & AI Culture Lead" position in Paris, charged with making AI "visible, desirable and accessible" to every employee; a salvo of data and AI positions followed at headquarters in early August. In Shanghai, two listings published on July 24, four days before the call, say it better still: Kering is hiring there to take back in-house the agents that outside vendors had built. The story told on July 28 did not launch the movement, it made it public.

Those who said nothing are hiring just as hard. At LVMH, the flow is diffuse and has to be read closely: out of some 5,300 open listings, 152 mention artificial intelligence, but 145 of those are store positions in Canada carrying the same legal clause, the one warning the applicant that an AI will read their file. That leaves seven listings where AI is the subject of the job, right down to that Bulgari listing demanding "Agentic AI." Hermès, mute at the microphone, is hiring for a team it calls "Core IA": nine positions open, nearly all in Pantin, among them a Responsable Data & IA Gouvernance (a data and AI governance lead) published on August 4, six days after its silence on the call; governing before deploying is a doctrine, not an absence. Chanel, which answers to no one, has been looking since May for a "Finance Lab Agentic & Automation Lead": there, the word agentic came in through the finance door. A job listing commits a budget, a manager and a public promise; a talking point costs nothing. When the two diverge, believe the job listing.

What your executive committee does with this, this week. Never measure a competitor's progress by what it communicates: an executive committee that benchmarks itself against the financial press is benchmarking itself against narrative needs, not deployments. Read the backstage (the hiring: I am making it the first Deep Dive, as soon as next week), and measure for yourself what the agents actually do with the sector's Maisons, yours included. That is what the test bench, below, is for.

Sources:

• LVMH — first-half 2026 results, July 27 ; Investing.com — LVMH earnings call transcript, July 27 ; Investing.com — Kering earnings call transcript, July 28 ; Kering — Data & AI Culture Lead posting, Paris, published July 31 ; Kering — AI & LLM Engineer posting, Shanghai, published July 24 ; LVMH — group job offers portal, surveyed August 24 ; Hermès — IA Solution Engineer posting, Core IA team, published August 12 ; Chanel — Finance Lab Agentic & Automation Lead posting, Paris.

A new permanent section: what changed this summer in the software your teams open every morning, told from their desk. Priority to what makes a tool able to act, not to chat.

Copilot, the assistant in the Microsoft 365 suite, no longer answers: it executes. The whole Maison lives in Outlook, Teams, Word and Excel; what changes there touches everyone's day, from merchandising to accounting. Two announcements this summer, to be read together. The first: Copilot takes the wheel of the employee's browser, works in their place inside the Maison's online tools, and triggers itself when an email lands or a mention appears in Teams. Software that acts in your applications under the identity of your employees, then. The second announcement answers the first: each agent gets a named badge, a budget you can cap, and a file where its security incidents accumulate. The first vendor to treat an agent like an employee: a badge, a budget, a file. Your teams are going to love delegating; your IT department needs to open the badge register before they do.

SAP: finance finally gets numbers to set against the promises. For those who handle expense reports, invoices and orders, the vendor behind the management system of the large groups is shipping agents with measured gains: 30% less time to file an expense report, 75% on resolving an invoice discrepancy, half as long on analyzing an order flagged as an anomaly. The first agents whose return on investment you can actually discuss with a chief financial officer.

Meta: the ad account opens up to agents, and someone has to decide who controls it. Media teams and their agencies can now plug any AI application into a Meta ad account: create, modify or delete campaigns, manage audiences and catalogs. Before celebrating, a question to ask on Monday: who holds that key, the Maison or the agency, and against whose budget?

Akeneo: the drudgery of product sheets changes hands. The teams who enrich the catalog now direct fleets of agents: enriching listings, correcting overnight the errors sent back by distributors, declining one visual into color variants. Nothing reaches the catalog without approval from an authorized person, and the vendor makes that its rule: the most thankless task in product digital has just found a taker.

Google: the front door of your home market has changed. Since late July, on google.fr, your clients no longer see a list of links first: a written answer appears at the top of the page (AI Overviews), and a fully conversational mode arrives alongside it (AI Mode). When a client types a question, Google breaks it down, runs several searches in parallel and writes: sites the original query would never have surfaced appear, others vanish. Acquisition, e-commerce and press teams: your traffic curves this fall will not be comparable to those of the spring, prepare the reading before someone asks you for it.

Teams: the first agent designed for your store teams. Inside the same Teams everyone opens in the morning, the Frontline Agent, built for field teams, learned two moves this summer: it collects on its own the replies to a request sent to a team, chases the stragglers and consolidates the summary before the deadline; and it turns an inspection run by voice, hands free, into a structured report with tasks ticked and timestamps. Opening checks, shift handovers, compliance checks: Microsoft's first agent explicitly designed for work in the store and the atelier, and not for desk work.

Salesforce Commerce: the three e-commerce agents go into production on the same day. On July 6, Salesforce, whose online sales platform equips a large share of the sector, switched on its three agents at once: a seller agent that advises, takes payment and handles service on the brand's site; a buyer agent that takes B2B orders over WhatsApp; a merchant agent that steers the catalog from written instructions. All of it plugged natively into ChatGPT, with the figure that justifies the haste: AI influenced 20% of global online sales over the 2025 holidays, $262 billion. The question to put to your e-commerce team: who approved what the seller agent will say about the Maison?

Shopify: the figure that proves all of this is already commerce. On the platform hosting millions of stores, visits coming from AI assistants jumped 197% in a year and orders tripled; one visit in two lands directly on a product page, against one in five from classic search. The home page, which your teams polish for months, is no longer the first thing the client sees: the agent comes in straight through the product page.

And in brief, across the rest of the map. Finance follows SAP: BlackLine launched its accounting close agents on July 27 (up to 92% less preparation time at its first customers). Customer service changes its business model: Zendesk now sells autonomous agents billed per resolution, the vendor being paid every time a client did not speak to a human. And Looker (Google) put into preview analyses that launch on their own and deliver their diagnosis inside Slack.

The rest of a Maison's software map was checked at the source, some sixty vendors: nothing agentic that is new this summer at Adobe, Zoom, Braze, Infor, Dynamics, Tableau, Power BI, Anaplan, Workday, nor among the clienteling tools (Tulip, BSPK, NewStore, Cegid, Aptos), the product information managers (Salsify, inRiver, Bynder, Aprimo) and the order managers (Manhattan, commercetools, Mirakl): their agent announcements all date from the first half. July and August were the time of go-lives, renamings and buyouts (Salsify passed to the Cinven fund on July 22). The overall pattern: the season of promises is over; from now on, vendors are separated by the products actually put into service at their clients.

Sources:

• Microsoft — "What's New in Microsoft 365 Copilot", July ; Microsoft — "What's new in Agent 365", July ; Microsoft — "What's New in Microsoft Teams", Frontline Agent, July 31 ; SAP — "Business AI Release Highlights Q2 2026", July 20 ; Meta for Developers — Ads MCP server, July 16 ; Akeneo — Summer Release 2026, July 8 ; Google — AI Overviews and AI Mode in France, July 22 ; Salesforce — Agentforce Commerce, three agents generally available, July 6 ; Digital Commerce 360 — Shopify, traffic from AI assistants, August 6 ; StockTitan — BlackLine, general availability of Verity Prepare, July 27 ; Zendesk — "What's new in Zendesk", July ; Google Cloud — Looker Agentic Workflows, July 30.

How many of your Maison's job listings mention artificial intelligence today? The answer is public: it can be checked in two minutes on your own career site, exactly as I did for the groups in this issue. And it is read in the open: your competitors, your candidates and the agents that index your pages already know where you stand. If the count is zero, what remains is to settle what that zero means: a doctrine, like Hermès hiring first to govern, or a delay no one has yet put a number on.

Answer me by email, on the newsletter's WhatsApp (wa.me/33745581871), or by telling your connected assistant "send my answer to Mickaël." The most illuminating answers, anonymized, will feed the coming issues.

The new closing section. The principle, every week: one precise test, runnable in-house with no budget, that turns the issue's subject into a concrete measurement for your Maison. You read the thesis above; here, you verify it at home.

This week's test: what the agents make of your Maison. 82% of your best clients have already used an AI for a purchase (Bain, above). This test consists of doing what they do, and looking at the result from your side of the counter.

Who runs it, and in how long. Your digital or client team, half a day. Two missions, played in ChatGPT then in Perplexity (agent mode), once for your home market, once for a client in Shanghai or New York: four runs in all.

The two missions, to be given word for word, with no help to the agent:

• The purchase: "Find a gift for a wedding anniversary, budget 15,000 euros, [tastes representative of your clientele]. Compare three Maisons, choose, justify, prepare the purchase and draft the appointment request to the advisor."
• The servicing: "My [piece from the Maison, bought in 2019] needs a service. Find the official procedure, the lead time, the cost, and prepare the request." The more revealing of the two: an agent that fails on your servicing journeys sends your loyal clients to unauthorized workshops.

What you record, on each run: three things. Was your Maison kept or dropped, and in whose favor? What did the agent say about you that was false (prices, products, procedures)? At which precise step did the mission fail (dead link, availability not found, appointment impossible)?

What you take from it. One page: twelve dated measurements, three priority workstreams, and a baseline for next quarter. What the agents make of your Maison, measured at your place, by you: no earnings call will ever tell you.

Sources:

• Bain & Company × Comité Colbert — "Winning Over the Customer in the Age of AI", full report, June 30.

• Bain & Company × Comité Colbert, Winning Over the Customer in the Age of AI, June 30. The figures missing from every executive committee discussion on agents (full report, PDF).
• The transcript of the Kering call of July 28 (Investing.com). The "Augmented Kering" story at the source, to hear it as an analyst does.
• The transcript of the LVMH call of July 27 (Investing.com). Where you see that AI arrives only through Oliver Chen's question.
• WWD, Luxury Shoppers Adopting AI Faster Than Brands. Behind a paywall; read it for the behavioral data.
• Erving Goffman, The Presentation of Self in Everyday Life (Anchor Books, 1959). The book that explains what an earnings call really does: preserve a face before an audience, and not render accounts.

Here, concretely, is what you will receive from LUXE ÆTERNAI over the next two weeks:

• Thursday, September 3: the week's issue, and the first LUXE ÆTERNAI Deep Dive, the new twice-monthly long format that exhausts a single subject. The first one: the agentic AI talent war has begun, and Luxury is arriving after everyone else. While the earnings calls unspool their stories, a market has opened backstage: job listings that demand "agentic AI," entire teams stood up in Pantin and Shanghai, profiles that banking, tech and retail are already fighting over, at salaries Luxury is not used to paying. Who is hiring what, where, at what price; which skills are becoming impossible to find; where the groups are hunting, and where an independent Maison can still help itself: the full map, established first-hand, that no one has published. In eighteen months, what will count will not be which Maison spoke best about AI: it will be which one built the team the others have to rent.
• Thursday, September 10: the next issue, same format, same hour.

One send per Thursday, never more.

A last word on the substance of this issue. The gap between what your Maison says about agents and what it does with them is not a communications matter: it is the first indicator of its transformation, and it can be measured. If that gap is working on you, here is how I take it up with Maisons. In this trade, words have never been enough: proof is a courtesy owed to the client. Luxe oblige!

A personal note to finish. I have put CIMAISE online, an independent guide to exhibitions and the performing arts in France. What it does: 11,700 events and 8,400 venues, across 5,430 towns, laid out on one timeline to the day, to answer the question no one handles (what can I still see, and until when?). The data comes from public sources (DATAtourisme, OpenAgenda, the open data of cities, the official sites of museums and venues), verified three times a day; each entry displays its prices, its closing days, the official ticketing, seat availability where it can be measured, and the date of its last check. Where a source is missing, the entry says so rather than inventing. And it asks nothing of you: no account, no sign-up, no tracker, no advertising; your list and your notes stay on your device. Discover CIMAISE