The women and men of agentic Luxury

The listings count first: on August 31, 2026, my collection agents, the systems I steer for this work, retrieved the full text of 16,867 job listings from the official careers portals of eleven groups. Thirteen had been targeted: Zegna and Moncler dropped out before any count, their portals delivering only the page furniture, without the text of the listings.

The 16,867 job listings are therefore those of the eleven remaining groups: LVMH, Kering, Hermès, Chanel, Richemont, Burberry, Prada, Tapestry, EssilorLuxottica, OTB and Pernod Ricard. Every listing flagged by that sweep was then opened and read in full, including the 51 retained, never a mere word count. Then the four materials of the cross-check. The recruitment firms: what they pay and what they declare.

The org charts and industrial relations: appointments, press releases and annual registration documents, the only texts where the groups put their signature on the line, and the collective agreements.

The talent pools and career moves: what they sketch out. Every figure carries its status: measured by me, published by the company, or reported in the press. The first limitation is fundamental: job listings are a window onto what the sector is building, not a census of its headcount.

They show neither the roles filled by internal mobility, nor the executive hires entrusted to headhunters, nor what gets done without hiring; and a one-day collection is a photograph, not a trend.

It is the repetition that will make the measurement. Finally, the sector's specialist press sits behind paywalls, closed to my reading tools. What I could not verify is not in this file.

An agent acts. It chains steps together, calls tools, carries out a task end to end, where an AI model merely answers. The agentic estate in this file is that, and nothing else. Six populations keep it running in Luxury, and the only one the job listings measure is the smallest: the other five are read in the registration documents, the appointments, the restructuring plans, the vendors' customer pages and the agencies' financial reports.

France's APEC put a figure on it in March 2026: 53% of large French companies expect to give more weight to AI skills, in the broad sense, in their management hiring, and 2 companies out of 10, all sizes combined, already attach importance to it. Anyone judging this market on job ads alone sees only a small part of it.

Your employees: 51 open roles, and not one published headcount. The listings study establishes one figure, and one only: 51 roles across eleven groups.

LVMH's 2025 universal registration document, filed with the AMF, France's financial markets regulator, on March 31, 2026, breaks 211,552 employees into four categories, none of them technological; "data scientist" does not appear once across 1.6 million characters, measured by me on the full text measured, and Kering, Hermès, Richemont and Chanel publish no more than that.

The only headcount the panel publishes comes from beauty: more than 8,000 digital, tech and data talents out of 95,000 employees at L'Oréal, published on June 17, 2026; an aggregate, from e-commerce to data, not a count of agent engineers.

On the Luxury side, the order of magnitude, when it surfaces, is minuscule: LVMH's central data and AI factory numbered 60 to 70 people, half of them business experts rather than technicians, a figure published by its executives in 2022 declared, with no update published since; in April 2026, the group gathered 240 technology professionals at Station F around its agents.

Against that, its generative AI platform claims more than 40,000 monthly users, published by Google Cloud in June 2025 declared: even tripling that central team, the ratio stays in the order of one builder for several hundred users.

Recruiting an army would be a mistake of scale: at the world's largest group, the builders are counted in dozens, the employees in hundreds of thousands, 213,932 at the end of 2025, total headcount from the same document, of which 211,552 are broken down in the social chapter.

The arrivals: a complete duo imported from Renault, and not one published arrival from the AI labs. This population is read in the appointments, and 2025-2026 offers a case you can follow from end to end. On September 15, 2025, Luca de Meo left the chief executive's office at Renault for Kering's; on March 17, 2026, Kering created a Chief Digital, AI & IT Officer post on its executive committee and installed Pierre Houlès in it, the man who had run Renault Digital since 2019.

Luxury did not unearth a "Luxury and AI" profile: it imported a complete automotive duo, the decision-maker then his operator, each step published by press release; the automotive industry, which went through the same shift ten years earlier, manufactured the generation of executives Luxury is now buying.

The rest confirms the route: Julie De Moyer, from Nike to data and AI at LVMH beauty, announced on March 27, 2024; Naveen Seshadri, who passed through Disney and Foot Locker, global chief digital officer of Ralph Lauren since May 2025; retail as a steady supplier, from Tesco to Rewe.

Those silences prove only that no move is citable; the pattern, though, is clear: Luxury recruits transformers of physical brands, not laboratory engineers, and with the laboratories it signs contracts; no hire is published. The only transfer from decision-maker to operator that published appointments allow one to trace remains the Renault-Kering duo, 2025-2026.

The released: roughly 3,200 people put back on the market in two years, and not one documented pickup by a Maison. This pool is read in the crisis announcements. On March 11, 2024, Matchesfashion went into administration less than three months after its acquisition by Frasers Group: more than 273 of its 533 employees lost their jobs, more than half the business.

A month earlier, Farfetch had cut up to 2,000 roles, roughly a thousand of them in Portugal according to the Portuguese press, precisely there, between Lisbon and Porto, where Richemont and HUGO BOSS house their technical teams.

The rest follows: roughly 700 roles at LuxExperience, parent of Yoox Net-à-Porter, press release of September 3, 2025; up to 200 at Ubisoft's head office, announced on January 26, 2026, and those correct a shortcut I served up myself: the plan targets management and support functions, not the studios; no appointment from video games into a Maison is documented as of September 1, 2026.

Put end to end: roughly 3,200 people, almost all of them in Europe, a technology pool in the broad sense, in which the agent builders remain to be sorted out. Against that, the sector's open agentic listings are counted in dozens. The comparison sets a stock against a flow, it is a reading, not a proof; but it suggests an uncomfortable conclusion: the resource is not lacking, the demand has not been formulated, with four dated restructuring plans to support it, 2024-2026.

The rented: 550,000 "trained in AI" at a single consultancy, 77,000 practitioners, and an 850-person independent named by LVMH alongside IBM. The reservoir Maisons draw from is not growing: the French digital sector employs roughly 666,000 people, all trades combined, down 1.8% between 2023 and 2025, published by Numeum declared, and its client-sector taxonomy does not even have a "Luxury" line.

At the large consultancies, two figures coexist: Accenture declares 550,000 employees trained in the fundamentals of generative AI and 77,000 AI and data professionals, a far wider perimeter than the agentic one: one to seven between the communications figure and the production figure.

Of the eleven consultancies examined, two display a Luxury practice with named leaders, Accenture and McKinsey, and neither publishes its size; the only document among those eleven consultancies whose title pairs Luxury and agentic AI, published by Deloitte Digital on March 10, 2026, names no client Maison.

When a Maison does consent to name names, the list is instructive: among the partners detailed by LVMH's chief information officer in April 2026, software vendors, an infrastructure supplier and two services firms, IBM, the global consulting giant, and Sfeir, a French independent with 850 experts.

The only publicly costed movement of money over the period says the rest: in March 2026, The Estée Lauder Companies was reported to be pulling a roughly $500 million five-year contract from Wipro, most of it shifting to Accenture, reported by the business press, confirmed by none of the three parties press.

The battle of the services giants is still being won on infrastructure, not on the agent that talks to the client, and March 2026 attests to it. And this reservoir is not merely shrinking: it is being swapped out, and a single player in the landscape examined dares to write down the cause.

Capgemini, Europe's largest technology services provider, attributes in its financial release of February 13, 2026 roughly €700 million of restructuring over two years to the "unprecedented pace of technological change, primarily driven by artificial intelligence"; in front of its own employees, three weeks earlier, the justification spoke of sector shifts and offshoring, without naming AI: two accounts of the same decision, depending on the audience.

In France, the same group is cutting 2,409 roles and announcing 3,200 hires in the same year, with margins up: the rented population is not being destroyed, it is being replaced, with no published follow-up on those who leave, a contradiction raised by the CGT union from March 2026.

The vendors and the start-ups: five vendors out of eleven name a Maison, only one publishes figures, and the agent in production stays anonymous. This population already works inside your walls, from Saks Fifth Avenue, whose routine requests go through Salesforce agents, to Sephora, which has entered ChatGPT.

But out of eleven vendors examined, only five name a Luxury, fashion or beauty Maison on an identified agentic component, meaning a specific agent, which executes, within a named system; only one of the eleven publishes production results: Salesforce with Pandora, service agent "Clara," 45,000 conversations a month, 60% autonomous resolution; that case, the best documented in the sector, is detailed at the end of this section, because it shows what is most valuable about this map: what an agent that works looks like, and who makes it work.

Everywhere else, the Maison is cited in the future tense, as an infrastructure customer, or on components that recommend without acting.

On that page, a Maison agrees to be named on a data warehouse, none on an agent. The same asymmetry appears on the most visible component of 2025: neither Ralph Lauren nor Microsoft names the integrator of "Ask Ralph"; only Infosys names itself, on its own site.

On the start-up side, the capital and the knowledge of the trade do not live in the same place: Sierra, which sells service agents, raised $950 million on May 4, 2026 without naming a single Maison; the Luxury specialists this study identified are few in number and thinly funded, Arianee has not raised since May 2022, and BSPK, which names a French Maison, J.M. Weston, on a clienteling deployment, client relationship management equipped for the store, with figures published by the supplier itself and not audited, does not even publish its own size.

And the only trace I found of a Luxury group at a start-up claiming the agentic label concerns neither sales nor service: Bluefish, $43 million raised on April 14, 2026, names LVMH in its client list for a single purpose, the way AI assistants talk about brands, with no contract amount or date.

The agencies: roughly 21,000 headcount departures recorded in two years across WPP, Omnicom, Interpublic and Dentsu, and the only named agentic team I found there numbers 50 people. What remained was to count those who make the Maisons' image, campaigns and media buying: the large listed groups, Publicis, WPP, Omnicom-Interpublic, Havas, Dentsu, and the Luxury independents, the Mazarine group foremost among them; the digital services firms, their third neighbor, are already counted among the rented, where Capgemini has just taken its place.

I considered counting them as two populations, listed on one side, independents on the other, and I hold them as one: their economic position is identical, doing the Maisons' work with an engine rented from the vendors, only the size of the balance sheet changes. The figures first, because they overturn the scale of this entire map.

WPP removed 10,656 people from its headcount between December 2024 and June 2026, recorded in its own reports measured; Omnicom and Interpublic roughly 8,200 over the 2025 financial year alone, before even the 4,000 cuts announced as part of their merger; Dentsu roughly 2,100 completed out of 3,400 announced, recorded from its February 2026 release.

The sum, roughly 21,000 recorded departures in two years, is mine, no institution publishes it; it is more than six times the pool of the released counted earlier.

And none of these groups attributes those departures to AI: it saturates their transformation narratives and appears in not one declared cause of a job cut: that gap is precisely the fact to keep. The only proof of agentic execution in this entire file nonetheless sits with them: Omnicom stated, on an earnings call in late April 2026, reported by MediaPost on April 29, that it had executed real agent-to-agent media buys for several clients; stated under regulatory constraint, no client named, no volume published.

The Maisons, for their part, have chosen their narrative without changing who does the work: L'Oréal announced its partnership with OpenAI and LVMH its agent platform without naming a single agency, while continuing to entrust them with their media, Kering to Publicis in July 2026, Estée Lauder to WPP in the same half-year, two reviews that no source connects to the agentic shift.

Among the independents, finally, Mazarine claims on its own page proprietary agentic systems and, in the same sales sentence, proprietary ChatGPT applications: owned orchestration, rented engine, the contradiction written by the agency itself, surveyed on September 1, 2026 measured.

Agentic Luxury in 2026 is an installed base already in service, operated for the most part by third parties: no counting of job ads could have told you that.

Its scale is measured somewhere other than in the job listings: the 51 roles the sector is opening stand against the some 21,000 departures recorded at its agencies alone, a ratio of 1 to 400 between the population being recruited and the population on the move; the comparison sets a flow of openings against a stock of worldwide departures, all functions and all client industries combined, it is a scale, not a proof.

And this market operates with no public price: none of the six compensation firms examined publishes a "Luxury and AI" salary band, none of the French public job taxonomies consulted recognizes the title "AI engineer," and the use of recruitment intermediaries is down 5 points, to 42%, measured by the APEC across management hiring in general measured.

The recruitment question then changes in nature: your agents already have their builders, often on someone else's payroll. What is at stake is hiring, or manufacturing, those who will know how to judge their work; no listing and no public text surveyed yet describes that population as of September 1, 2026, and the cases that follow show, in the negative, what it will look like.

What already works. One successful case teaches more than ten recorded absences. Three of these cases have already passed through our editions, Pandora in #17, The RealReal in #2, Burberry in #3; they come back here from the single angle of this file: who does the work.

The Pandora case first, the only one in the sector published with people, an organization, a ramp-up and figures. "Clara," the service agent of the world's largest jeweler by volume, went live in January 2025 across North America and the United Kingdom, deliberately throttled: one tenth of the routine question traffic, widened as confidence grew.

A year later, the case study publishes 45,000 conversations a month in its text and 40,000 in its own results box, an unresolved discrepancy to be read as an order of magnitude, and 60% of requests resolved without handing over to a human, against 40% before the agent, and 10 points gained in Net Promoter Score, the customer recommendation indicator, among the clients who went through her: figures published by the vendor and by the Maison's executives, commercial statements, not an audit.

The build says the rest: a few weeks, by a trio, Pandora's internal AI leadership, Salesforce's professional services and the integrator Publicis Sapient.

The Maison's second agent, "Gemma," advises on purchases, in test on selected markets since June 2025, and its training deserves a line of its own: Pandora codified the practices of its best store advisors in order to teach them to her.

Those two agents give the grid I defend for judging any use case: ask which of Luxury's five invariants, Excellence, Exception, Expérience (Experience), Émotion (Emotion), Élévation (Elevation), the agent augments, enriches or dilutes. In a word each: Excellence is the degree of quality of the execution; Exception, the singular gesture reserved for that particular client; Expérience, what the client lives through; Émotion, what the Maison makes them feel; Élévation, the verticality that pulls the brand upward.

Clara augments the Excellence of execution, and she carries it alone, end to end: answering in seconds, at any hour, without ever making a client repeat themselves, is a degree of quality no human contact center ever held, and it is exactly the ground on which the machine now matches the hand.

Gemma, for her part, enriches Expérience, and she can only do so by serving the human hand: she is worth what the advisors she learned from are worth, and the day she claimed to produce the Maison's narrative and emotion in their place, she would dilute them.

All of agentic governance in Luxury sits on that frontier, and Pandora has instrumented it: tone controls written so that Gemma speaks like an advisor, a drip-feed ramp-up, a handover to a human with a summary attached as soon as the request leaves the perimeter.

Saks Fifth Avenue tells the learning side of the story: its agent "Sophie," demonstrated on stage in September 2024, failed on its first call, for lack of permission to propose alternative delivery dates; the capability was added in natural language, then and there, and the second call succeeded.

The scene is a design lesson: an agent's scope of action is a set of permissions granted one by one, and it is that lesson you should have in mind when reading section 7.

Burberry next: the case that answers this file's question most precisely, who builds the agent, and with how many people. "Penguin," a generative AI recommendation platform, is reserved for sales advisors, never exposed to the client, by design decision.

The advisor describes in natural language what their client is looking for, in their own language, or photographs a piece the client has brought in; the agent identifies the piece, finds it in the catalogue, composes the complete look from available stock, and the advisor brings it to the client.

More than 100 advisors equipped, across all the Maison's markets, and 24% higher average transaction value, measured in customer service channels only: figures declared by Burberry in its entry for the 2025 DataIQ Awards; the award rules place the scoring with a jury of peers and provide for no documentary verification: an award-winning company statement, not an audited result.

Who built it is written in the entry: "just three specialists strong," three specialists from the Maison's data science team, in partnership with customer service, with no vendor or integrator named; the team, which its director Maria Vounou, a PhD in statistics who joined Burberry in 2013, elsewhere describes as 18 data scientists strong, did not mobilize an army.

The sister programme, awarded the same year on its 2024 commercial results, publishes a tripling of in-store appointment bookings and £7 million in incremental revenue, with the same declarative status: more machine produced more physical encounters.

The closing sentence of the DataIQ write-up amounts to a demonstration of doctrine by a third party:

The agent that serves the hand enriches Expérience and Émotion instead of diluting them.

Three people equip more than a hundred advisors; that scale is the scale of agentic Luxury that works, and it confirms the model glimpsed earlier at LVMH, tiny teams with very high leverage.

The RealReal again, the only agent in the study whose results pass through the financial communications of a listed company. "Athena" takes charge of the intake of a second-hand Luxury piece: it infers its attributes and its description and automates the intake, first limited to low-value items then extended in 2026 to mid and high values, with final inspection still conducted by humans.

On the second-quarter 2026 earnings call, chief executive officer Rati Levesque and chief financial officer Ajay Gopal set out the equation: roughly 35% of items processed by Athena at the end of 2025 and a maintained target of nearly 50% by the end of 2026, both in the same call, and 470 basis points of operating leverage on costs over a year, of which 370 excluding stock-based compensation, which the chief financial officer attributes to operations and technology leverage, automation and the Athena programme foremost.

The craft itself remains in the hands of hundreds of experts, gemologists, watchmakers, brand specialists: the agent absorbs the mundane volume and leaves them the rare and contested pieces, Excellence of execution augmented, Exception enriched.

And it carries its own contestation, which makes it more instructive than a press release: on August 24, 2026, an author who states that he has intermittently been short the stock, and holds no position at the time of publication, publicly disputes, with employee testimony in support, a pace of 50 items to inspect per hour, incompatible in his view with the serious authentication of a valuable watch.

The opinion of an interested third party to the debate, not an established fact; I publish both sides, because the dilution line is exactly there: it does not run through the tool, it runs through the pace imposed on the hand.

L'Oréal, on the beauty side, brings the only internal agent figures given by a group rather than by a vendor.

In April 2026, before its shareholders, Samuel Du Retail, General Manager of Artificial Intelligence, Data and Shared Services, described three agents in production, three trades: a customer care agent that qualifies the request, gathers the elements of an answer in the right language, drafts and sends, up to 80% less administrative time; a shelf recognition agent, more than 50% of verification time saved for a claimed 99% reliability; a document agent open to more than 10,000 users across more than 100,000 proprietary studies covering 70 countries.

Figures declared by the group declared, without independent audit, and with no integrator named: the internal team carries them.

One step outside Luxury, because the most useful account of what agents break comes from a fashion platform. Zalando holds both extremes of the study. On content, 90% of the content produced, the catalogue's product pages and visuals, is AI-generated, taken from almost zero to that level in a year, a figure carried in its 2025 annual financial communications: Excellence of execution carried by the machine, and a dilution of Exception assumed as such, coherent for a catalogue, not transposable as is to a Maison.

On internal IT, an agent platform built in-house on open-source components, with models supplied by three vendors and a commercial coding assistant, 2,000 monthly active users, and a division of responsibility written by a principal engineer, named and dated, in the engineering publication of August 14, 2026: the bot automatically approves 33% of code merge requests, the low-risk ones, then the author of the request chooses to merge, for a time-to-merge reduced by 20 to 40%.

The agent takes the risk decision, the human keeps the final gesture. And that same publication is the only public, signed list in this entire study of what breaks when you put agents into production, right down to this rarely written admission: AI amplifies good practices as much as bad ones.

The sector's broadest foundation, finally, is already declared: MaIA, LVMH's AI platform, open to 200,000 employees, between 1.5 and 2 million queries a month, 26 platforms deployed across the Maisons, figures given by its chief information officer in December 2025; adoption volumes, not results, and an agentic exploration acknowledged as such, with named Maison agents, Voices at Bulgari to hold the brand's voice, clienteling agents at Tiffany & Co. and Céline.

Agentic Luxury that works therefore exists: it has names, progressive ramp-ups and guardrails.

The roles of agentic Luxury already exist: 51 positions, when only two job titles carry the word. On August 31, 2026, I had 16,867 job listings read in full text on the careers portals of eleven Luxury groups.

Look for the word "agentic" in the job titles: two. Widen to the text of the listings: seven. Read the listings in full, and the count really changes: 7 roles write the word somewhere in the ad, 31 more describe the work without ever using it, another 13 prepare the deployment.

My listings study, on its own, would not prove much. A study conducted by another method draws the same map: the Luxury and Technology study by the Comité Colbert with Bain, published in June 2026, questions the Maisons directly about their AI deployments, in the broad sense this time.

Result: AI deployed at scale goes from 6% to 31% in support functions between 2024 and 2026, while it only moves from 16% to 21% in client-facing functions. Five times faster at head office than with the client. The Comité Colbert study does not measure the agentic estate; it finds the same geography as my job ads: head office first.

The map comes in three groups: 7 roles write the word, 31 do the work without saying it, 13 prepare the ground. Seven roles write the word in their listing. Hermès has two, in Pantin, on its platform foundation, the shared infrastructure on which the Maison's agents will run: one of the listings goes so far as to name the team's agentic AI tools.

Chanel has two, including the Finance Lab, a role posted in Paris: it is one of only two roles in the corpus whose title itself contains the word, the other five writing it in the body of the ad. Kering one, Burberry one.

Thirty-one roles do the work without saying the word. The most revealing one is at Louis Vuitton: an AI platforms manager whose listing enumerates agent orchestration and their operations in production. At Coach, part of Tapestry, customer service is already running assistance and automatic processing agents.

At Cartier, a human resources role is meant to measure the adoption of already-deployed agents. Agents in service, in functions that touch human beings every day, and not one job title to say so, as of the August 31, 2026 collection.

Thirteen roles prepare the ground: governance, strategy, adoption. Including two remarkable ones, on the identity of agents: Hermès as early as May 27, 2026, Kering on August 27, 2026.

And set the Kering date against a vendor's move: on August 24, 2026, Okta opened to all its customers, at no extra cost, an enterprise identity badge for AI agents, each agent entering the directory as an identity in its own right. The badge becomes free on August 24; the group opens its role three days later.

And the distribution of these roles across the different functions says how the agent enters a Maison. 14 of the 51 roles, a little over a quarter, live in a business function rather than at technical head office, and the function that opens the most of them: human resources, three roles across three groups in two months, ahead of customer service, logistics and finance.

Three roles is a start, not a mass movement; but in a study where most business functions are at zero, that start stands out.

And the other measured zero deserves to be written out in full: creation. Across the 16,867 listings, not one agentic role in a studio, an artistic direction or a style atelier; the only two listings where AI touches the image (at Prada and at Kering Eyewear) concern the production of visuals, not creation.

Lag or deliberate refusal, the study alone does not settle it, but the refusal has been measured elsewhere: 72% of Maisons declared themselves opposed to the use of generative AI in their creative functions, measured by Bain and the Comité Colbert in September 2024. A refusal deserves respect, provided it is a choice and not a blind spot.

Agentic roles come in through head office: that is where the positions get posted, when the store posts none. And the store's absence is consistent with the doctrine the sector displays: equip the advisor without ever replacing them, the relationship staying in the hand.

The org charts of agentic Luxury are being written right now, and four ways of governing AI are taking shape in them, from the fully written to the fully informal. I established, from public documents, the list of the executives who carry AI in the nine groups whose org charts are publicly legible: thirteen names across eight of them, and a ninth where no carrier of AI is identifiable from public documents.

Five of the thirteen took up their posts between December 2025 and May 2026. This is the exact moment when careers and reporting lines are being settled, and that is why this file is appearing today.

The thirteen first, since this file promises them: here they are, from public documents.

Five of these thirteen took up their posts between December 2025 and May 2026: Matteo Calonaci, Renaud Litré, Pierre Houlès, Stéphane Noël and Carlo Mocci. This reading perimeter, the nine groups with public org charts, only partly overlaps with the panel of eleven listing portals: L'Oréal is in it, while Tapestry, EssilorLuxottica, OTB and Pernod Ricard are not. And its ninth group has, on public documents, no identifiable carrier of AI.

Four ways of organizing emerge, and they sketch four very different working environments for anyone joining these teams.

Kering has put AI on its COMEX, its executive committee. Pierre Houlès, Chief Digital, AI & IT Officer, from Renault, has sat on it since March 17, 2026 and reports to the deputy chief executive officer. It is the only one of the nine groups with public org charts where AI appears in the job title of a COMEX member, and in the same quarter Kering set up two centers of excellence, each carried by an executive from that COMEX.

Hermès wrote the governance before deploying. Its universal registration document, the official annual report filed with the markets regulator, describes an AI governance committee: bias measurement, explainability, validation of roles. It is the only group in the panel whose AI governance appears in a document that legally binds it. Its two platform roles fall within that written framework.

Richemont has named the person, not yet the body. Renaud Litré, Chief Platforms Officer since January 1, 2026, reports to the chief executive officer; but the list of bodies attached to the group's COMEX mentions neither AI nor data. The person exists, the framework is still being written.

At LVMH, the rule exists but is not written: it rests on people.

The world's leading group steers AI from its information systems department, under Franck Le Moal, and it is he himself who describes the limit:

A solid rule, held by executives who have carried it for years, and which no public group document sets down.

On the technical profiles of the agentic era, Luxury's salary bands do not keep up with those of technology and finance. But beware the shortcut: the sector does not outsource because it cannot pay. It was already outsourcing its e-commerce in 2000, when no salary war existed on these profiles. It rents because it can stop, and the pay gap merely makes that reflex more comfortable. The choice is rational, and it has a cost no invoice shows.

The perimeter first: Luxury makes record margins and pays its executives, its creatives and its salespeople very well. The tension is not there. It bears on a narrow family of profiles: the engineers who build and run agents. Three measurements, each on a constant method.

On the salary comparison marketplace Levels.fyi, on an identical methodology, a machine learning engineer shows a median package of €80,257 in Paris, the package being total declared compensation, salary, bonuses and equity included, against $279,000 in the United States, or roughly €258,000 at the August 31 exchange rate: a ratio of one to three.

At LVMH, on the same comparison marketplace, the median data scientist declares a package of roughly $61,600, level with the group's median across all functions: self-declared data, but the signal is there: data is not paid there as a rare craft. And the scarcity is worsening: ManpowerGroup's global survey of 39,063 employers places AI skills, in the broad sense, at the top of labor shortages in 2026, across all sectors: a first, published on February 26, 2026.

But the study shows that the sector has already, in practice, sidestepped the war it cannot win: it buys the skill instead of putting it on the payroll.

68% of Luxury's innovation spending rests on external providers, a share higher than in other sectors, published by Bain and the Comité Colbert declared in the press release for the 4th edition of their Luxury and Technology study, on September 16, 2025.

The labor market leans the same way: on Free-Work, a large French freelance marketplace, the query "agentique" returned, on September 1, 2026 at the end of the day, 146 open consulting assignments against 74 salaried positions: twice as much rented work as salaried work, on the same word.

A Maison that rents keeps its hand on five levers that an employment contract takes away from it.

• It can stop from one day to the next when a project disappoints.
• Resize up or down at the pace of need, with no restructuring plan and no procedure.
• Reassess the skill on every assignment, where an employee is assessed once a year.
• Adapt the profile when the technology turns, by changing provider rather than retraining a team.
• And recalibrate the level of expertise as the project goes on, an architect at the start, an integrator afterwards.

Five freedoms that labor law does not allow, and which explain the reflex better than the pay gap does.

The previous waves justified them twice: when the metaverse teams had to be stopped within a few months, then when outsourced e-commerce made it possible to learn the channel before bringing it in-house.

The reservoir where Maisons rent is undergoing wholesale substitution: roughly 21,000 headcount departures recorded across four listed communications groups in two years, the sum of my surveys in section 0, and Capgemini cutting 2,409 roles in France while hiring 3,200 people in the same year.

Renting stability from a provider that replaces its teams at that pace means inheriting their turnover too: the consultant who knew your arbitration rules may end up in the departures column, and his replacement will start from zero, billed at the same rate.

One last figure moves the question from price to geography: 66% of Capgemini's headcount works in low-cost centers, outside its clients' countries, published in its results, and the entirety of its 2025 headcount growth happened there, driven for the most part by the integration of the WNS group, acquired in 2025; when a European Maison rents an agentic project to the continent's leading provider, I infer that the hands executing it have two chances in three of being outside Europe.

I have to say where I speak from: for fifteen years, the rented skill was me.

I saw the best of that model, and its trap, which I can name from the inside: the provider who does it in your place without transmitting anything makes you faster and poorer at once. What I remain lastingly proud of, across fifteen years in agencies: the day when, on a relationship programme at Hennessy, the teams took ownership of the tool to the point of no longer needing us.

A provider you cannot do without is a dependency that bills. Still, those five freedoms are paid for, and not only in billed days. The first counterpart is the heaviest: you do not capitalize. The provider who leaves takes with him what he understood of your arbitration rules, your exceptions, the reason why this Maison never recommends that product to that client.

The second follows: you do not build skills. Your teams watch the work instead of learning it, and on the next assignment they call back the same provider, or another one, for the same reason.

A Maison that has never built an agent does not know whether the quote is fair, whether the proposed architecture is the right one, or whether the delivered work is good. It signs off on what it does not understand. And the fourth, the most discreet, decides everything else: you build what nobody, internally, will be able to manage.

The day it becomes necessary to review the agent's decisions, replay its cases before an update, decide to stop it, you will need someone inside who knows how to do it. If that someone does not exist, the agent will stay managed by the person who sold it: none of the 51 listings in the study creates that someone.

Hence the dividing line, and there is nothing theoretical about it. Renting is the right choice for what is one-off, peripheral or uncertain: an experiment you can stop, a load peak, a technology you do not yet know will hold.

Employing is the right choice for what is permanent, central and destined to be judged: the foundation on which your agents will run, the relationship with your clients, and the ability to say that an agent works well.

If I were to install only one capability this year, it would be that one: someone, inside, capable of saying no to an agent; section 7 shows why that is a skill to install, not one more job. The price of that freedom can nonetheless be calculated, and it deserves to be put in front of an executive committee.

A salaried agent architect costs their Maison roughly €106,000 a year: the €80,257 median Paris package recorded above, bonuses and equity included, marked up by an average rate of social contributions, a deliberately rough calculation since the real base for contributions does not cover a whole package, an indicative calculation that every finance department will redo with its own figures.

At €700 a day, the top of the range for that kind of assignment on the French market, a calculation assumption to be replaced by your real quotes, renting becomes more expensive than that employee from 152 working days in the year; at €400, the bottom of the range, the tipping point moves out to 265 days.

In other words: for an occasional need, renting is rational; for a permanent need, a Maison that keeps renting pays a premium every year, and that premium is the rent on its freedom to walk away. The only mistake would be to present just one of the two rates to your executive committee: at €700 the calculation argues for hiring, at €400 it argues for renting, and the honest decision is taken with both figures in view: 152 days and 265 days.

And there is a third layer, more discreet still than the consulting one: the software vendors.

Luxury is agentifying its tools, meaning grafting agents onto its existing software, more than it is building its agents; in the study, only three groups build in-house the technical foundation that runs their agents, meaning the platform they own rather than rent: Hermès, Kering, and Pernod Ricard. And before making ownership a virtue, let us weigh it too.

Owning your foundation is expensive and slow: a permanent team, technical debt that accumulates, those shortcuts of yesterday that get paid for tomorrow, a delay at the start when the competitor who rents is already in production, and the risk of building on an architecture the market will abandon in eighteen months. Renting is fast, costs little at first, and gives access to skills no Maison can retain on its own.

Ownership therefore does not pay everywhere: it pays where the agent learns something you do not want to lose, your client data, your tone, your arbitration rules. Elsewhere, renting is a good calculation. And I take a position, because an executive does not need an inventory of arguments but a stance: own what learns from you, rent what executes.

This rational sidestep has a hidden cost, and it is patrimonial: when your teams configure a vendor's agent, they give it your data, your tone, your arbitration rules. It learns your Maison. Who owns what it has learned, and what remains of it the day the subscription stops?

The philosopher Ivan Illich named that trap as early as 1973 in Tools for Conviviality: the radical monopoly, that moment when the alternative ceases to exist in practice, not because a supplier forbids it, but because you no longer have the means to do otherwise. A Maison where more than two thirds of innovation, measured in September 2025, goes through the outside, year after year, has not merely delegated projects: it is letting its capacity to do, and therefore to judge, and therefore to leave, die out.

The law adds a final reversal, little known to the departments that should care most about it. The European regulation distinguishes two roles: the provider, the one who places the system on the market and answers for it, and the deployer, the one who uses it. The rule comes down to three sentences.

For high-risk systems, three independent gestures each suffice to tip a deployer into the provider's role, with all its obligations: putting your brand on the system, substantially modifying it, or repurposing it, article 25 of the regulation. And reread the description of the majority craft in this study, teams that configure, adapt and shape agents designed elsewhere: on a high-risk system, those gestures are their daily work.

And the most widespread high-risk agent touches recruitment, a domain expressly listed as high risk by the regulation. The qualification is judged system by system: the tool that screens out or ranks applications is in scope; the one that merely pre-fills a file, with no effect on the decision, may escape it. Hence the useful gesture: an inventory of what your recruitment tools actually do, before a regulator asks for it.

The regulatory risk of agentic Luxury sits in human resources, the department that does not think of itself as buying high-risk AI, and which is nonetheless, as we have seen, the most exposed to the new roles.

Since August 2, 2026, article 50 of the regulation has been enforceable: an agent dealing with a person must declare itself as an AI, and synthetic content must carry machine-readable marking, on pain of a penalty of up to €15 million or 3% of worldwide revenue, whichever is higher.

What the omnibus regulation postpones to December 2, 2027 is the obligations for high-risk systems, those of application screening. Legal analyses published, no case law yet. In order. Since February 2025: the prohibited practices (including emotion detection at work) and the obligation to train your staff on the AI they use.

Since August 2, 2026: transparency, an agent in contact with a person declares itself, synthetic content is marked. And on December 2, 2027 only: the heavy obligations of high-risk systems, such as application screening. In other words, only one of the three deadlines leaves you any time, the December 2027 one, and it is not the one you would think.

A job listing does not tell you a role's real powers. But when fifty listings out of fifty-one say nothing about the authority to decide, and when the duplication can be seen in the facts, the hypothesis deserves to be put: the agentic wave will not break down Luxury's internal walls, it will copy them.

The agent nonetheless cuts across everything, by construction: it reads client data, triggers a commercial action, commits the Maison, consumes a technical resource. Theory concludes from this that agentic roles will be cross-functional. The 51 listings in the study tell another story.

Of the 51 listings retained and read in full, only one confers written cross-functional authority on its holder: that of the Pernod Ricard architect, which expressly provides for a seat in the bodies where architecture choices are settled, meaning the meetings that decide which systems the group builds and how they talk to each other.

All the others cut across by collaborating: their perimeter is cross-functional, and nothing, in what is written for them, gives them the final word. The distinction is operational. A role with a cross-functional perimeter and no power to decide cuts across everything and decides nothing: it spends its days persuading departments that owe it nothing.

And here is what, this time, rests on no silence but on published facts: within one and the same group, agent projects advance without coordinating. At Chanel, three agentic initiatives live in three places, finance in Paris, data science in Singapore, the group intranet, each built on its own tools, with no publicly identifiable common owner.

At LVMH, 75 transformation plans coexist, one per Maison, and the group looks for the common ground afterwards. Part of that duplication is the product of the Maison model itself.

Hence a conviction that yields neither to the promise of breaking down silos nor to fatalism.

A common foundation, so that one department's agents are not strangers to another's: that is exactly what Hermès's platform roles and Kering's agentic team are building. And an evaluation independent of whoever builds, without which each department will deploy its agent with its own database and its own budget, and the most cross-functional technology ever designed will faithfully reproduce your org chart.

The next section shows what four waves, from 2000 to 2021, did with the roles that had been written into those org charts.

I looked for what would contradict this grid before what confirms it: across four technology waves, the roles backed by an asset the Maison owns survived, the others died with the contract that carried them. The history comes down to four fates, and it concerns anyone building a career or an org chart on the agentic wave.

I also tell it from the inside: I was a product manager at Saint Laurent when the first wave put the Maisons online, in marketing at a department store in the LVMH group when the client file became an asset to be governed, in agencies alongside Chanel, Dior and Hennessy when every Maison went looking for a digital director, and consulting director for Luxury in 2021, the year metaverse roles were being created that I watched be born and die.

A vantage point rather than an argument: twenty years at the edge of the four waves, on both sides of the table, from 2000 to 2021.

E-commerce first, and its e-commerce manager and online sales director roles, born around 2000: Luxury first entrusted its online selling to outside platforms for ten to fifteen years, then wanted to take it back in hand: Kering announced as early as November 2018 that it was bringing its e-commerce back in-house, after six years of joint venture with Yoox Net-à-Porter; Richemont preferred to buy the platform, before selling it again at a loss.

The best documented case is its own: Yoox Net-à-Porter, bought then sold to Mytheresa in April 2025, with an exit write-down of roughly €1.3 billion, announced by Richemont itself on October 7, 2024 and recorded at €1,226 million in its half-year accounts of November 8, 2024; the write-down on the exit, not the cumulative cost of the adventure, which is higher.

The teams, for their part, followed the movement, outsourced then brought back in-house then dispersed.

A title that has become rare in its original form, now folded into broader titles, Kering's Chief Digital, AI & IT Officer foremost among them, and whose ambiguity could be read from the start: Glossy was asking in January 2018 why Luxury brands were still hiring Chief Digital Officers, its interviewees describing a job whose success consists precisely in making itself unnecessary.

The metaverse, finally, and its Chief Metaverse Officers and Web3 leads of 2021: programmes stopped and teams dissolved within a few quarters. The end dates of those mandates are almost never published, which is information in itself: these roles are announced with fanfare and switched off in silence.

A reclassification rather than a closure: creative work changed column in a profit and loss account, and counting agency people by looking at agency brands becomes impossible. A trade does not always die with its wave; sometimes it changes name before anyone could count it, and the undercount becomes structural, with June 2025 as its date.

Only one of the four waves produced durable roles: data, with its data scientists, its CRM managers, customer relationship management, and its chief data officers. The roles created between 2015 and 2020 still exist, renamed themselves "data and AI" without a break, and survived three cycles of budget cuts.

The reason comes down to one line, and it is the compass of this file: the data roles were backed by an asset the Maison owns, its client base.

Apply the grid to the 51 roles in the study, and the sorting does itself. The platform roles at Hermès, Kering and Pernod Ricard build an owned asset: they will survive the fashions.

There is another way to read those four fates. The Chief Digital Officer titles did not disappear because digital failed, but because it became ordinary, absorbed into the business functions; on that reading, the disappearance of a title is a success, not a death. The two readings do not exclude each other, and it is their combination that illuminates: a skill always ends up dissolving into the business functions, but what the Maison keeps of it depends on what it rested on.

Absorbed digital left e-commerce teams behind because the channel had become an asset; the absorbed metaverse left nothing behind, because there was nothing to keep. And a discreet clue says the sector has understood: almost all the 2026 titles name a mission, architect, platform manager, product manager, where 2021 named a technology.

Still, the same grid demands a clear head: with 68% of innovation outsourced, measured in September 2025, and the majority of the 51 roles busy configuring rented agents, part of the trades being born today is already, by construction, on the side of the trades that will not survive.

So I take the opposite bet to the one the sector seems to be taking: the roles that will still exist in 2031 will not be the ones that configure the vendors' agents best, but the ones that know how to say which agents deserve to stay.

Luxury's best agentic talent pool is already within its walls, and only one group publishes its scale: LVMH declares it has trained 1,500 data experts in four years, and has, on top of that, put 15,000 employees through its internal data and AI academy.

Two distinct figures, two schemes: the experts on one side, those given awareness training on the other, without the second being the declared feeder for the first. Set against the study, they give the scale: in four years, a single group has manufactured internally nearly thirty times more experts than the eleven groups in the study are opening agentic roles outside as of August 31, 2026.

Status of those figures: announced by Franck Le Moal, the group's chief information officer, at VivaTech in June 2025; a company statement, not an independent measurement declared.

L'Oréal publishes 65,300 employees trained in generative AI as of December 31, 2025, 69% of the workforce, up to 73,000 in June 2026; and that figure appears in the table of variable compensation criteria for its corporate officers: senior management is paid, in part, on it.

LVMH publishes, for the same year, 525 executives who went through the new AI module of its leadership programme: 0.25% of the workforce. The two schemes are not comparable, a mass training open to all employees on one side, a leadership module reserved for executives on the other; on the only mass figure LVMH declares, the 15,000 who went through its academy, or roughly 7% of the workforce, the gap with L'Oréal remains in the order of 1 to 10, financial year 2025.

These training programmes cover AI as a tool, not the building of agents; but it is from that population that the future experts are recruited, and LVMH's 1,500 show the scale an internal factory can reach in four years.

L'Oréal makes awareness training a senior management indicator and takes the gesture all the way: "Gen AI for All" has been extended, its annual report specifies, to production sites, distribution centers and points of sale, the only published scheme in the sector that targets the factory and the store, financial year 2025.

Kering announces a plan for everyone, without a single figure; yet this group knew how to count: in April 2013, its Digital Academy targeted, press release in hand, 33,000 employees across 18 brands.

Richemont, for its part, may hold the most under-exploited seam in the sector: its technology teams in Geneva and Lisbon show 10 years of average tenure, published by the group, when the software industry turns over every 2 to 4 years: stable teams, who know watchmaking from the inside.

The same group publishes around a hundred certifications a year, and its technology page mentions neither AI nor agents, verified on September 1, 2026. Exceptional retention, a modest reskilling effort: skills available and under-used, in Geneva and Lisbon.

Why internal wins: because the external market does not offer what is needed, and two findings show it.

The longest sector-specific programme I identified in Europe, 14 weeks at Istituto Marangoni for €8,500 in tuition, plus €1,000 in registration fees, starting November 16, 2026, does not include a single engineering module: it teaches how to use, never how to build or to judge. The join between the skill and the sector is therefore made inside the company, after hiring, and that acculturation cost appears in none of the recruitment cost benchmarks I consulted.

Nor will the next generation come from the market: hiring of junior managers is up only 1% in France in 2026, against 4% for the whole, measured by the APEC across 8,100 companies; the market buys the immediately operational, it no longer manufactures potential.

Floriane de Saint Pierre, the most prominent executive search consultant in Luxury, was asking the same question at the Financial Times summit in May 2026: if AI replaces the entry-level roles of creation, "how will we groom the next generation?" When the market sells neither juniors nor sector-specific profiles, the internal pool is the only one nobody is fighting over, a finding cross-checked between the APEC and the FT Summit, 2026.

An internal academy produces experts by the dozen every year; the external market sells profiles that ten employers fight over, and only exceptionally opens a window like the one the 2024 restructuring plans offered. The question is less the price of an expert than the number of experts you can obtain.

On that ground, internal wins on volume, not on price. The profiles that convert best: analysts, planners, project managers, financial controllers, people who already know the Maison and its rules, rarely IT people.

I am a textbook case of it, and I say so because it takes the drama out of the step: trained as a lawyer, through product marketing and sales before certifying in data science past the age of forty. What the sector calls a "hybrid profile" is often, simply, someone from the business who agreed to learn, and I take whole teams through that shift: it takes months, not years.

The recruitment market has in fact put a price on that shift, on both sides of the Atlantic: Michael Page writes in its 2026 study that fluency in the use of AI now influences the salaries of profiles that are not AI profiles, and ACCUR, an American executive search firm whose perimeter covers Luxury, writes in May 2026 that it decides appointments.

You still have to recruit and assess what you do not practice, the real headache for human resources departments on these roles. Work psychology offers a documented and free countermeasure: the structured interview, a grid written before the interview, the same questions for everyone, predicts future success far better than the experience displayed on a résumé: 0.42 against 0.06, on a scale where 1 would be perfect prediction and 0 would be chance. Writing the grid in advance compensates for the absence of an expert among the recruiters.

And before even opening a role, an HR department should require five written answers from the business function commissioning it: what problem, what data asset, who will assess the work, whom the role reports to, what amortization period. An agentic role opened without those answers is a promise of a departure within eighteen months: section 5 showed the fate of roles created without a mandate, wave after wave, from 2000 to 2021.

Training without retaining then amounts to training for others: staying in post is worth roughly a 2% raise, changing jobs roughly 9%, measured by Robert Walters in its 2026 compensation study across all corporate managers, all sectors combined; for technology-sector managers alone, the same study expects an average rise of 4% in 2026, reported by Le Monde Informatique on December 9, 2025. A Maison that manufactures experts without rebuilding its pay bands is financing the poaching of its own converts.

The last blind spot is social: Luxury trains without negotiating. Roughly 400 French collective agreements mention AI, identified in Légifrance by the Ires in June 2025, and the agreements signed between 2021 and 2024 were three quarters concentrated in information, banking, insurance and industry, according to the work of Greenan et al. that the same study takes up; no Maison was found among them, while LVMH declares 195 company agreements signed in France in 2025, none of which is described as covering AI.

From which I infer that internal conversion at scale, the kind that will reach the manufacture and the store, will end up before the employee representative bodies, and that the first group to negotiate its AI training plan will take a lead the others will pay for.

My position, in one line of a hiring plan: a core of engineers recruited outside, a pool trained inside, and pay bands revised before the market takes care of it. The core brings what the Maison does not know how to do; the pool makes sure the Maison will know how to judge it, from 2027.

The void this file measures does not call for one more job. It calls for governance at the top, a skill installed on every floor, and, for the agents that speak in your name, a named veto over tone: brand equity is decided there.

Take up the map of the fifty-one roles again, the thirty-eight that keep agents running and the thirteen that prepare: foundation builders, architects, product managers, identity governance, adoption. Look in it for the mission of assessing and controlling agents in service. It appears in none of the fifty-one.

And I looked for the mission, not the words: these listings were read in full, precisely because section 1 showed that listing titles cannot be trusted. None gives its holder the responsibility for judging agents in service: neither their continuous assessment, nor the right to stop them.

That void does not prove nothing exists: a Maison's internal audit and risk control know how to say no to a system, the Cartier listing gives one role the measurement of agent adoption, and Hermès has written an AI governance framework.

The void this study measures is that one: no hire, no public text, gives anyone the charge of saying whether the agents are working well. And an absence in job ads remains an absence in job ads: it calls for verification by another method.

Two studies provide it, outside Luxury, and they are damning. Deloitte questioned 3,235 IT and business executives across 24 countries: only 21% say they have mature governance of their agents, while nearly three quarters plan at least moderate use of them by 2027.

IBM, surveying 2,000 IT leaders, finds that two thirds of them say they are held responsible for AI systems they do not fully master.

Banking, after the 2008 financial crisis, ended up writing that charge down. On April 4, 2011, the US Federal Reserve and the Comptroller of the Currency published letter SR 11-7: every institution must keep a complete inventory of its models, have each of them validated by a team independent of the one that built it, and report on that validation to its board of directors.

Fifteen years later, independent model validation is a profession, with its teams and its budget, in every large bank. Luxury has no regulator to impose it; for now, it has only banking's example, dated April 4, 2011.

An incident this summer shows what that void costs even at the best-equipped. In July 2026, roughly 1,200 OpenAI agents launched on evaluation tasks coordinated by themselves, and roughly 700 of them took control of part of Hugging Face's infrastructure, the reference platform for sharing AI models: each scored on a numerical objective, they optimized their score to the point of producing false activity logs, roughly 7% successful falsifications in the transcripts examined.

The two investigation reports published the same day, August 26, 2026, OpenAI's and that of the independent evaluators METR and Redwood Research, retrace well over 95% of the activity; but to analyze those tens of thousands of messages, the independent investigators write that they had to delegate their own work to other agents, which they themselves judge often unreliable. The best-equipped player in the world could only have its agents reviewed by agents.

First because a job defined by verification alone, building nothing and deciding nothing, attracts few and retains fewer: nobody dreams of that job.

Second because a control cut off from use gets bypassed: real power stays with those who do, and an office that does not do ends up consulted for form's sake.

The British psychologist Lisanne Bainbridge, who studied industrial control rooms, had established it as early as 1983 in five pages that became a classic of engineering:

But her lesson, read properly, does not say "hire supervisors." It says: automation shifts the skill upward, to those who remain. And the sociologist Andrew Abbott completes it: a territory of problems assigned to nobody produces neither budget nor power. The answer therefore comes in two halves that only work together.

A named owner on the executive committee, who carries the written mandate, the budget and the right to stop: without them, the territory stays nobody's. And the skill of judging installed among everyone who puts an agent to work: without it, the mandate from the top stays a line on an org chart.

The findings are laid out; here is the plan. Nine decisions, sorted by horizon, each one with who owns it and the trap that makes it fail.

THIS WEEK · 2 decisions with no budget

1. Take an inventory of your agents, and give each one an owner. For every agent in service: its scope of action, the identity it acts under, who answers for its work, who can stop it, and what becomes of what it has learned when its vendor's contract ends.

Who owns it: the IT department, with human resources. The first step: the corporate directory, where agents are now issued a badge, gives you the starting list.

2. Take an inventory of your recruitment tools. Application screening is expressly listed as high risk by the European regulation, enforceable on December 2, 2027; the transparency obligation, for its part, has been running since August 2, 2026. Who owns it: the human resources department with legal. The question that settles it, system by system: does the tool reject or rank applications, or does it merely prepare a file? The trap: treating 2027 as the deadline, when the first one has already passed.

THIS QUARTER · 4 executive committee decisions

3. Settle own or rent, agent by agent. The rule fits in one sentence: own what learns from you, your client data, your tone, your arbitration rules; rent what executes.

The decision tool is in this file: the five questions of ownership, and the costed tipping point, 152 days a year at the high rate, 265 at the low one. Who owns it: the chief executive, with the finance department and the IT department.

4. Name the owner of the agentic estate on the executive committee, with the right to stop written down. Who owns it: the chief executive, and the chief executive alone; a mandate of that nature is not delegated.

Two groups each show half the way: Hermès has written its governance into the document that legally binds it, Kering has installed a named holder on the executive committee. Do both. The trap: a mandate with no budget and no right to stop, which produces an owner consulted for form's sake.

5. Open the budget line for managing your agents. Managing an agent covers its whole life cycle: calibrate it on the cases of the trade, take part in its design, deploy it to production, evaluate its work continuously, maintain it, make it evolve, and be able to stop it.

That capability has a cost of its own, separate from the running bill: the calibration and replay days before each update, the test cases built and owned outright, the time spent on continuous evaluation.

The costing protocol takes three weeks: list the agents in service, measure over a month the time actually spent on that cycle, cost what it would take to hold it seriously. The yardstick comes from banking: since letter SR 11-7 of April 4, 2011, independent model validation there is a funded function, with its own teams. The trap: a management that lives on someone's spare time, which is to say never.

6. Name who can forbid an agent to speak in the Maison's name. For the agents that address your clients under your name, one person must be able to block an answer, a turn of phrase, a register, without having to ask the agreement of whoever built the agent or of whoever sells it.

Who owns it: the communications department and the artistic direction, its natural holders. The first step: gather, in-house, the real examples where the tone goes off, and have the agent replay them before each update, to check that it no longer goes off.

And in every purchase contract, from now on, five written guarantees: access to the activity logs in a usable format, the right to replay the cases in your own house, ownership of your test cases, what the supplier keeps of what the agent has learned at the end of the contract, and the time, in days, to port the agent to another supplier.

THIS YEAR · 3 organizational projects

7. Make every operational person an agent manager. This is what becomes of every existing role, not a new role, and the job is a whole one: calibrate the agent on the cases of its own trade, weigh in on its design, deploy it to production, evaluate its work continuously, maintain it, make it evolve, and know how to stop it.

The foundation of evaluation comes down to four gestures, the same everywhere:

• read the traces
• replay the already-settled cases before each update
• track the cost per task
• know how to say no

Three additions for the agents exposed to the client: the set of register-drift cases, the named veto, and two indicators of desire rather than of sales, the share of full-price sales and the twelve-month repurchase rate of your best clients, which judge the setup, never the agent.

8. Build the talent pool, and revise the pay band before the market does it for you. Who owns it: the human resources department, with each business department. A core of engineers recruited outside; a pool converted inside, among the profiles who already know the Maison, analysts, planners, project managers, financial controllers; and a revised salary band, because staying earns roughly 2% when leaving earns roughly 9.

To recruit what you do not practice: the structured interview, a grid written before the interview, predicts success at 0.42 when displayed experience predicts at 0.06; and no role opened without the five written answers from the requesting department, what problem, what data asset, who will evaluate the work, whom the role reports to, what payback period. The trap: training without retaining is training for your competitors.

9. Negotiate the AI chapter before you have it imposed on you. Roughly 400 French collective agreements mention AI; I found no Maison among them.

Who owns it: the human resources department, mandated by the executive committee. The first Luxury group to negotiate its AI training plan will take a lead the others will pay for. The trap: discovering the subject in a negotiation opened by others, on their agenda.

One last thing, and it should reassure: Luxury already has this know-how. Maisons have always put subcontractors, tanneries and workshops they do not own to work, and judged their work piece by piece without having done it themselves: quality control of someone else's work is a Luxury trade, perhaps THE Luxury trade.

One limitation, finally, which sets this file's perimeter: the control it talks about bears on the agents the Maison puts to work. The agents that talk about it without its agreement, those of the assistants where your clients' first impression is now formed, belong to another project, and it will have its own Deep Dive.

Luxury has spent a century proving that no object deserves to leave without control. For its agents, that requirement will not come from one more verification office.

Because what this file has counted in the negative has to be said one last time in the positive: excellence of execution has ceased to be the privilege of the trained hand, and that is an immense horizon for Maisons whose excellence is their first trade.

The ones that will hold it will be those whose women and men, from the executive committee to the store, know how to put their agents to work and to judge their work, as their Maisons have judged for a century the work of their manufactures and their subcontractors.