The Window.

Where luxury stands.

August 2026.

The market found a floor and has not yet found an engine.

Worldwide luxury spending reached €1,443 billion in 2025. Personal luxury goods, the part most independent houses live in, dipped to €358 billion, down 2% at current exchange rates and up 1% at constant rates. Bain assigns a 70% probability to its base case of 2 to 4% growth in 2026, reaching €365 to €373 billion.

The first quarter of 2026 was between minus 5% and minus 3%. Around 60% of players are already outperforming their own first quarter of 2025, and the wide performance gap that defined last year is beginning to close as the strongest performers cool and the laggards recover.

This is not a recovery narrative. It is a market that has stopped falling, in which the distance between a good house and a poor one is closing from both ends.

The base underneath the market is changing shape.

BCG and Altagamma, in the twelfth edition of their True-Luxury study, describe four structural rebalances now under way.

• 01

  The share of aspirational consumers in the revenue mix moves from roughly 70% in 2016 to roughly 50% in the next cycle.

• 02

  One-time buyers fall from around 60% of the customer base to around 40%, as the sector shifts from acquisition to retention.

• 03

  Traditional personal luxury categories rebalance from around 80% of the mix toward 60%, as spending moves toward lifestyle and experience.

• 04

  Cross-border sales, historically inflated by tourism, fall from around 50% to below 30%, as domestic wealth becomes the structural engine.

Every one of those four movements favours a house that knows its clients by name over a house that depends on footfall. This is the most consequential paragraph in this section for a reader running an atelier of forty people.

The top tier is the engine, and it has been for a decade.

Top-tier clients now account for 24% of luxury spend, up from 14% ten years ago, while representing 0.1% of luxury consumers. Their spending is uncorrelated to the macro cycle. Aspirational consumers have supplied almost all of the market’s volatility, and this year their fall is finally slowing.

Average annual spend in the top-tier group runs at roughly €360,000, with a minimum threshold above €50,000. Including luxury mobility and wellness, their share of total spend reaches 37%.

The ultra-wealthy population, corrected.

There were 556,850 individuals worldwide with a net worth above thirty million dollars at the end of 2025, holding $63.8 trillion between them. The population grew 14.4% in a single year, the strongest expansion since 2017, and has added more than 120,000 people in two years. Altrata forecasts 746,570 individuals and $85 trillion by 2030.

These are the clients that matter to a house of this size. They are also the clients most independent houses cannot currently identify, reach, or serve at the level they now expect.

What the client buys.

When BCG asked more than ten thousand True-Luxury respondents to rank the drivers behind their most recent purchase, the top three were design and aesthetics, craftsmanship and quality of execution, and timelessness. Logo visibility ranked last across every product category.

Price rejection now runs through every tier: 70% of consumers walked away from a purchase in the past year because they judged the price unjustified. More than half of them stayed inside the brand or the sector, moving to a different piece or a competitor. They were not lost. They were redirected.

The client moved first.

This is the section that changed most between editions, and it is the reason the paper carries a new subtitle.

Trust is where the change shows.

87% of luxury consumers use generative AI weekly. 40% use it daily. Of those, around 80% already use it to research luxury: asking for recommendations, then comparing options, across both personal and experiential categories.

Adoption was predictable. What was not predictable is where these tools now sit in the client’s hierarchy of trust. On BCG’s net trust measure, the engines score 29 percentage points, placing fourth among the main information sources for luxury. That places them level with traditional web search and almost level with word of mouth. They are trusted at roughly twice the level of social media and influencers. Only a house’s own website ranks higher, at 42 points.

The reason consumers give is that they perceive these tools as more neutral than social channels, with the absence of sponsored content acting as the differentiator. Whether that perception is accurate is a separate question, and I take it up in Section XII. What matters commercially is that the perception exists and is already shaping purchases.

Where the client accepts AI, and where the client does not.

BCG tested seven concrete use cases across client-facing domains, measuring awareness, brand perception and purchase intent. 62% of consumers now expect AI involvement across key luxury use cases. 83% say they would perceive a brand positively or unchanged on discovering it uses AI. 77% say they would keep buying from such a brand.

The acceptance is not uniform, and the pattern of it is the single most useful finding in this paper.

• Essentially no resistance.

  After-sales support, experiential clienteling, product design, and AI-generated product and still-life imagery for digital channels.

• Real resistance.

  AI-generated advertising campaigns, and personalised clienteling messages presented as coming from a named client advisor. The resistance concentrates almost entirely among older consumers in Western markets.

Read that division again. The client accepts the machine in the workshop, in the service that follows the sale, and in the room where the relationship is deepened. The client objects where the machine impersonates a person or authors the house’s public voice.

The thesis of the first edition was that AI belongs to intimacy rather than scale. I wrote it as a conviction. The client has now drawn the same line, in a survey of ten thousand people, and drawn it in the same place.

And the industry is behind.

60% of fashion and luxury companies remain at an emerging stage of AI maturity. A separate BCG study across a thousand companies put roughly 57% of luxury firms in the laggard category, with little implementation.

For a large group, being behind is a budget problem. For an independent house it is a window, and the window is the subject of this paper.

The second behaviour the headlines still miss

The due-diligence query before the meeting.

The private client runs a due-diligence query before the meeting. They ask who this house is, what people say about it, and what the risks of working with it are. Whatever the engine says is what the client believes when they walk in.

Most independent houses have never tested this. When they do, they find the engines describe them as smaller than they are, less established than they are, and occasionally carrying coverage from years ago presented as current. In Section XIV I set out what the testing found.

The remedy is a quarterly ritual: probe the engines, document the answers, correct the record. It is the cheapest hour of strategic work available to a founder, and almost nobody in this segment has it in the calendar.

What the big groups have already built.

You cannot compete on infrastructure. You need to know what is there, because the floor rises for everyone.

The case study that has now outgrown the case study

Brunello Cucinelli, Solomei AI, and Callimacus.

Brunello Cucinelli, a family-controlled Italian house, built its own research company, Solomei AI, and through it a platform called Callimacus: a page-less, menu-less presentation layer in which an orchestra of agents interprets a visitor’s intent and composes the experience in real time. It launched publicly in January 2026 in Italy, the United States and the United Kingdom.

The team was seven people. Mathematicians, an engineer, an artist, a philosopher. Three years of work.

What has happened since is the part that matters for this paper. Time spent on the site has doubled, and conversions followed. More than forty inbound enquiries arrived from other industries, most of them outside fashion and luxury entirely. And in July 2026 Salesforce agreed to invest in Solomei AI, joining the Cucinelli family holding company as a shareholder.

An independent house built an instrument shaped to its own philosophy, and a company with a market capitalisation in the hundreds of billions bought into it. That is the clearest available proof that the signature instrument argument in Section VI is a commercial argument, and that it survives contact with a balance sheet.

The technology horizon.

Where things stand, August 2026.

Correction

The first edition said inference cost was down roughly tenfold in eighteen months, and elsewhere that build costs had fallen roughly tenfold in twenty-four months. The measured figure for inference is far larger and the measured figure for delivered build cost is smaller and less precise. I have separated the two claims in this edition; they were being used interchangeably and they are not the same thing.

The worked example of a shiny thing failing

In-chat checkout, September 2025 to March 2026.

In September 2025, OpenAI and Stripe launched in-chat checkout with a great deal of noise about the end of the shopping page. It was retired in March 2026, after roughly thirty merchants had integrated and near-zero sales. The protocol survives as infrastructure. The consumer product does not.

Meanwhile the discovery layer, the part nobody demonstrated on a stage, is where 80% of GenAI-literate luxury consumers already are. I include this because a technology horizon that only lists what worked is a sales document, and because the pattern will repeat: the demonstrable thing gets the budget and the boring thing gets the client.

Next twelve to twenty-four months, through mid-2028

• Agentic discovery consolidates.

  Well before agentic checkout does. Position in the answer layer becomes a budget line under whatever name the industry settles on.

• Voice and ambient interfaces enter luxury retail.

  Smart eyewear with multimodal capability has been under active evaluation at the three big groups for some time, and evaluation is a thing large groups are very good at.

• Digital twins and bespoke configurators.

  They normalise in watches, jewellery and fashion.

• AI-augmented client advisors.

  The operational lever, put to work inside the low-resistance domains BCG identified, and kept out of the others.

• Digital Product Passport preparation, ahead of obligation.

  The textiles delegated act is expected late 2027 with compliance realistically 2028 or 2029. The first mandatory passport is the battery passport in February 2027.

Three to five years, 2028 to 2031

• Humanoid robotics.

  They reach luxury hospitality back-of-house and high-touch manufacturing support. Treat the market forecasts in circulation, including the ones the first edition quoted, as vendor-sourced until an independent series exists.

• AI-designed materials.

  Driving fabric, metallurgy and composite innovation, following sportswear by two to four years.

• Private and local inference becomes standard.

  Making the sentence about client data never leaving the house technically true rather than rhetorical.

• The curated corpus overtakes the model.

  As the defensible asset. This was the first edition’s prediction and it is the one I would now make more strongly, not less.

Why the small house has the edge.

Ten to one hundred and fifty people. Five structural advantages the groups cannot buy.

A discipline most AI papers miss

Some knowledge should remain unwritten.

Not everything the house knows belongs in a model. The reasons a watchmaker still finishes bridges by hand on surfaces nobody will ever see are partly economic, partly aesthetic, and partly beyond either.

That knowledge is oral, embodied, and present only in the apprenticeship of the next generation. A model cannot hold the lineage. It can only support the conditions in which the lineage survives.

What to build. One signature instrument.

Four reasons this is the play

• 01

  The cost has collapsed, and the window closes as the category standardises. What is bespoke in 2026 is a template in 2028, priced higher and differentiating nothing.

• 02

  The clients who matter will pay for it. 0.1% of consumers now drive 24% of spend, and what they pay for is intimacy, theatre and time.

• 03

  It is defensible. Anyone buys a subscription. Nobody replicates your archive, your philosophy, or fifteen years of trust with a specific family.

• 04

  The economics beat every other marketing investment available to a house this size. A campaign shoot runs eighty to two hundred thousand euros and lives six months. A signature instrument runs forty to a hundred and fifty thousand euros for a full build, and lives three to five years.

The houses that will matter in 2030 are the ones with a product built on top of the technology that still feels, unmistakably, like them.

Budget and revenue.

Real numbers, with the weak ones removed.

What the industry is reporting.

41.2% of luxury companies are implementing generative AI in selected areas and 11.9% have embedded it in core functions. Executives rank artificial intelligence the single largest change ahead of them, at 31.7%, with materials and production innovation second at 22.6%. Customer experience and loyalty is the strongest perceived growth opportunity, at 28.6%.

Annual AI and data budget, by revenue band.

These are working figures from my own engagements rather than a published series, and they are presented as such.

• €1M to €5M revenue.

  €30K to €120K a year, 1.5 to 3% of revenue. One tier-one product at minimum viable scope, plus tier-two foundations.

• €5M to €20M.

  €120K to €400K, 1.5 to 2.5%. One full tier-one instrument, tier-two foundations, and the visibility work.

• €20M to €80M.

  €400K to €1.6M, 1.5 to 2.5%. Two tier-one instruments, full tier two, one tier-three initiated.

• €80M to €300M.

  €1.6M to €6M, 1.5 to 2.5%. Full tier one and two, two tier three, an in-house lead.

Two warnings on these numbers

• Data preparation.

  For a house of this size it runs at 40 to 60% of any real project, loaded into year one. The temptation is to skip the unglamorous data work and go straight to the visible product, and it is the single most expensive mistake available. Budget 150 to 200% of initial development for the five-year total.

• Break-even.

  One to three years on well-scoped projects. Whatever you build, instrument it, measure it, and kill it by month nine if it is not earning its place. MIT’s review of some three hundred public deployments found roughly 5% producing measurable impact on the profit and loss, and named the cause as organisational. I use that figure with a caveat, because it has been quoted far more often than it has been read and its method is thinner than the headline implies. I use it anyway, because it matches what I see.

The stack.

No religious wars. The best stack is the one your team can run.

Three principles that override the stack

• Vendor portability.

  Do not lock into a single model provider. Prices and capabilities will move every six months for the next three years. One chief technology officer, on my own file, quoted €180,000 to migrate a mid-sized luxury build off a single provider when prices changed. The abstraction layer would have cost €4,000 to install at the start.

• Data sovereignty.

  For Swiss, French and Italian houses serving private clients, either the client data does not leave the European Economic Area or you hold a written contractual chain of custody. There is no third option that survives a client asking the question.

• Human in the loop by default.

  Every outbound communication routes through a person until you have twelve months of clean logs. This is also where BCG’s resistance data lands: the client objects to the machine writing as a named advisor, and the approval step is what keeps that line intact.

Process, systems, governance.

What the regulation requires, as of this month.

The EU’s Digital Omnibus on AI was endorsed by Parliament on 16 June 2026 and approved by the Council on 29 June. It deferred the AI Act’s high-risk obligations to 2 December 2027 for standalone systems and 2 August 2028 for AI embedded in regulated products.

It did not defer Article 50. The transparency duties, including disclosing to a person that they are interacting with a machine and marking synthetic content, applied from 2 August 2026. New prohibitions follow on 2 December 2026. Fines for prohibited practices reach €35 million or 7% of worldwide turnover; transparency and high-risk breaches sit in a lower band.

The relief is real, and it is relief on the obligations an independent house was least likely to trigger. The duties that did arrive on schedule are precisely the ones a forty-person house breaches by accident, on a Tuesday, through a website assistant nobody remembered to label.

The six documents you need on paper.

• A House AI Charter.

  One page. What the tools may do, may not do, and what requires a signature.

• A Data Handling Map.

  What client data you hold, where it lives, what trains on it.

• A voice and aesthetics guide written for machines.

  Tone, paired examples of good and bad output, boundaries.

• The Refusals Register.

  What the house never does, regardless of price. Maintained by the founder with a senior craftsperson, signed annually. This is the deepest piece of brand work the technology requires, and most houses have never written it down.

• AI literacy training records.

  Required under Article 4, in force since February 2025.

• A risk register.

  Fabrication, data leak, voice drift, regulatory exposure, provider outage.

The four rituals that make governance real.

• 01

  A monthly output review of one hour, in which the founder reads a random sample of machine-assisted output across every channel. Skip it and the brand slips without anyone noticing.

• 02

  A quarterly council of ninety minutes across operations, marketing, craft, technology and legal, whose main job is killing projects that are not earning their keep.

• 03

  A quarterly probe of what each major engine says about the house. The cheapest hour on the calendar.

• 04

  An annual external reading by someone who asks the questions you have become too close to see.

The human layer.

Where the implementation fails. Not from technology. From grief.

What the founder is seeing

• Picture the senior salesperson who has spent eighteen years writing letters to her clients in her own voice, asked to let a machine draft them.

• The atelier head whose juniors used to come to him for research that now takes seconds.

• The marketing coordinator who learns that the part of her job she trained for is the part the machine does best.

None of these people is resisting the technology. They are mourning a part of themselves that the work used to hold.

And note where the client agrees with them. The one place consumers push back hardest is the machine writing as a named human being. Your salesperson’s objection is not sentimentality. It is a commercial instinct that the survey data confirms.

Four moves.

• 01

  Name the loss. The founder says it aloud, in a room, in person. An email does not do this work.

• 02

  Budget the forty to sixty hours. For a forty-person house that is two thousand to two thousand four hundred hours across year one. Put it in the plan or it will not happen.

• 03

  Offer new identities alongside the new tools. Salesperson to curator. Coordinator to editor and orchestrator. Junior researcher to judgment specialist.

• 04

  Let the team choose the first targets. Ask them which of their tasks is most tedious. The wins are then theirs, and trust follows from that. It has never once followed from a presentation.

SWOT for the independent house.

The risks nobody puts on the glossy slide.

A five-year foresight.

By 2028

• 01

  Agentic discovery is a normal part of the top of the funnel. Closing remains human-assisted above five thousand euros, and I would now say that threshold holds longer than vendors predict.

• 02

  Passport infrastructure is live and the first category obligations have landed. Watches and jewellery are still preparing.

• 03

  AI-assisted clienteling is table stakes for houses serving private clients, put to work inside the low-resistance domains and kept out of the rest.

• 04

  The AI Act’s high-risk rules bind from December 2027, and at least two or three groups have faced enforcement. The lesson will be loud.

By 2031

• 01

  The best-performing fifth of houses have a signature instrument that belongs to the brand, and does not sit beside it like a campaign.

• 02

  A handful of independent houses have outgrown mid-tier groups by serving five hundred to two thousand private clients with an intimacy that does not scale the old way.

• 03

  The proprietary corpus, client data, archival knowledge and craftsman notes, appears in investment memoranda as an asset with a number beside it.

• 04

  Private inference is standard, and the sentence about client data never leaving the house is a technical fact.

What the Report is measuring.

A research paper that only reads other people’s data is a summary. This section is the part of the argument being tested here.

The pilot.

Earlier this year I put the questions a wealthy buyer asks, the ones that come before a boutique visit rather than after, to ChatGPT, Gemini and Perplexity, and read what came back with twenty years in the industry behind the reading, and no scoring rubric. Three findings held across all three engines.

• 01

  All three deflected the buyer away from retail using near-identical depreciation figures, none of them sourced to anything. The same number, phrased three ways, presented as fact.

• 02

  A certified horological first was stated in full by one engine and softened by another, because the second had drawn on forum posts instead of the maison’s own record. A four-year development, adjudicated by a message board.

• 03

  One independent maison appeared in none of nine intent-based buying questions, while the same four or five rivals filled every answer. The house did not place low in those answers. It was not in them.

None of this is the engine behaving badly. It is the engine doing what it was built to do with the material available to it. The maison had a hundred years of craft and almost nothing a machine could read.

The Report.

is published by ELIA Atelier and measures how AI answer engines describe luxury houses. Edition one, published 27 August 2026, records 17,820 answers across 46 Swiss watch houses, nine buyer questions, four engines and three markets.

Through August 2026 those forty-six houses are read across ChatGPT, Gemini, Claude and Perplexity, against the buying questions their clients ask, in English. Edition one publishes at 07:00 CET.

The large visibility studies published this year cover software, retail and consumer categories, where the samples are easy and the buyers are obvious. The independents, the segment with the most craft per employee and the least machine-readable record, have been measured by nobody.

Why the answer layer behaves the way it does.

Two findings from the wider literature explain most of what the pilot surfaced. Research from the University of Toronto found that the overwhelming majority of AI answers cite third-party material rather than the brand’s own site. And a study of 1,094 buying categories tracked monthly in ChatGPT through the first half of 2026, covering more than fifty thousand brands, found that only 15.2% had a clear owner while 53.7% were unsettled, with no brand appearing in even three of five related questions. Once a brand did own a category with a comfortable margin, it held first place in more than nine of ten month-on-month comparisons.

The same study found that only 21% of the most-cited domains in a category were also the most-mentioned brand, and that the two signals correlate slightly negatively. Citation and recommendation are different things. Your client reads the answer; nobody reads the footnotes.

On whether the watch findings generalise

I expect they do. Nothing the pilot surfaced is specific to watches. The engines were not penalising horology; they were reaching for whatever material existed and finding forums, resale listings and press aggregation where the house’s own record should have been. A jewellery house, a perfumer, a leather atelier and an architectural practice share that shape exactly: deep craft, a founder who carries the story in person, a thin published record, and a category in which four or five large names dominate everything written online.

So I expect the pattern to repeat, and I am not going to claim it until it has been measured. The cohort is watchmaking, and I will not claim beyond it. Jewellery and perfumery are next in the programme.

I did not set out to build a visibility practice. I set out to read what the machines were saying about houses I have known for twenty years, and I did not like the answers.

Where the marketing budget should go.

One

The client already trusts this channel more than the incumbent one.

On BCG’s measure, the engines carry a net trust score of 29 points as a source of luxury information, placing fourth overall, level with traditional web search and close behind word of mouth. Only a house’s own website scores higher, at 42 points. These tools are trusted at roughly twice the level of social media and influencers.

Now ask what proportion of your marketing budget currently goes to the channel your client trusts least, and what proportion goes to the one now sitting level with search.

Two

Position compounds. A campaign does not.

A campaign shoot produces a six-month run of impressions and returns to zero. A trade-show stand produces three days of conversations and is taken down. Neither compounds, and both are budgeted annually as though the question were settled.

Position in the answer layer behaves differently. Across 1,094 buying categories tracked monthly through the first half of 2026, brands holding a category with a comfortable margin kept first place in more than nine of ten month-on-month comparisons. And 53.7% of categories had no owner at all: no brand appearing in even three of five related buying questions.

A category with no owner is not a category to advertise into. It is a category to take.

Three

Most of the work is not advertising, which is why it is affordable.

The overwhelming majority of AI answers cite third-party material rather than the brand’s own site, and brands are surfaced through someone else’s pages several times more often than through their own. The levers are the house’s published record, the third-party corroboration the engines weight above it, and the entity clarity that lets a machine work out that two spellings of the same name belong to the same house.

None of those is a media buy. All of them are editorial, and editorial is the one discipline a house of forty people can do better than a group of forty thousand.

What I am and am not arguing.

I am not arguing for zero media spend. I am arguing that a house allocating six figures to paid and influencer work and nothing to the answer layer has a budget shaped by 2019. A first serious allocation in this segment costs less than a single campaign shoot, and unlike the shoot it does not expire in September.

Two cautions, because a section like this one is where papers stop being honest. The trust premium consumers currently grant these tools rests partly on the absence of sponsored content, which is a temporary property of a young market. And nobody can guarantee a position: the engines change their sources and their behaviour, and any vendor selling a guaranteed position is selling something that does not exist.

What can be built is the material the engines read and the record they read it from. That survives whichever engine wins.

How ELIA Atelier works.

The practice accompanies independent houses through this transition. Three rooms, one principle: build only what serves the craft.

Two ways to begin

• The Business Assessment.

  An online reading of where the house sits today, returned as a structured editorial profile across five dimensions. Free, and useful whether or not you and I ever speak.

• The Blueprint.

  A two-week strategic reading of the house across voice, operations, client experience, team, and data and memory. It produces three priorities, a ninety-day map, and a five-year horizon.

A working conclusion.

The open question now

The open question now is a duller one, and it is entirely within the founder’s control.

The client has already gone looking. The question is whether anything the client finds was written by the house.

The smartphone did not end the atelier. This will not be different. The houses that get it right will still be here in 2040.

A note about who is writing this

ELIA Atelier is the AI visibility practice for the luxury industry, founded by Alexandre Olive and based in Genève, Switzerland. Twenty years inside the luxury industry, thirteen of them at executive level in high jewellery and watchmaking, and a decade since then accompanying the founders who run independent houses.

Alexandre Olive

Corrections register.

Every figure removed or amended between the first and second editions, with the reason. Listed so that anyone who cited the first edition can correct the record.

Sources.

How to cite this paper.