Lost in Translation

Luxury judges on detail. Machine translation is improving mainly in fluency, and fluency hides the missed detail.

On October 13, the BBC interviewed translators and researchers. François Yvon, research director at the CNRS, observes that a good system renders the French idiom "il pleut des cordes" (it's pouring ropes) into English, but wonders whether it will manage in Indonesian. Carol Bereuter, a medical translator, describes errors that are harder to spot because the text "sounds perfectly legitimate."

In Dublin, on October 10, Trish Van Bolderen, who chairs the executive committee of the Irish translators' association, summed up the effect before a trade union: fluency "can trick you into buying what's being presented," according to the Dublin InQuirer.

The anthropologist Edward T. Hall distinguished in 1976, in Beyond Culture, so-called high-context cultures, where meaning lies in what surrounds the words, from those where everything goes through the words. Luxury speaks in high context: the way to address a regular client formally, the name of a house canvas one does not translate. The psychologists Rolf Reber and Norbert Schwarz measured the effect of reading fluency in 1999: a sentence that is easier to read is more often judged true. So a brand has its translated texts reviewed by someone who knows it in the target language, and asks that person what is missing, not what is wrong.

The RealReal hands the search for pieces to an agent. On October 8, the Luxury resale platform opened Ask TRR to its 45 million members, according to the press release distributed by Google Cloud. The agent searches in plain language across more than one million one-of-a-kind pieces and suggests a selection; the release does not credit it with any purchase. The platform also plans recommendations on its product pages, according to Luxury Daily on October 12, which relays the press release without any figures from the trial phase.

The next step is easy to guess: agents mandated by their clients that will search, find and buy on their behalf. A platform would then need fewer pages for humans than entry points for machines: direct access to its catalog, the MCP protocol (Model Context Protocol), which plugs an agent into a tool, or A2A (Agent2Agent), which lets two agents talk to each other.

Yet I do not believe in resale without humans. The agent will serve the buyer who already knows what they want, the reference, the size, the price; the treasure hunter will keep browsing. The one who delegates no longer experiences the find, the piece sought for years and unearthed oneself, and the brand whose piece is being resold loses a moment when its name mattered to them.

A Bulgari job posting in Shanghai mentions agents for client advisors. A job posting published on October 10 in Shanghai, for a data and AI role, lists among its duties supporting "Agentic AI initiatives" with client advisors and client relations teams. Of 696 LVMH job postings published from October 6 to 12, which I collected on the 12th from LVMH's recruitment portal, five mention agents: one at Bulgari, four at Sephora, a retailer, including an internship to automate task workflows. Like the postings collected in Edition 20, they show a team getting trained; none describes an agent already in service.

1. London consults on agents until November 20. On October 8, the British data regulator (ICO) launched a six-week call for evidence, open also to companies that deploy agents (ICO); Richard Nevinson, its director of technology regulation, warns that "the fact AI agents act with autonomy is not an excuse for poor compliance," according to the ICO, as reported by UKAuthority. Why it matters: a brand whose agent answers British clients is concerned, and can respond before the deadline.

2. Workday lets the client set its agents action by action. On October 13, the software maker announced agents that work in the background on recruiting, then payroll and accounting (SiliconANGLE, CIO); the client sets three levels, from automatic action to human decision, with review in between, according to CFO Dive. Only recruiting is open, to early customers; payroll and accounting will follow in early 2027. Why it matters: if your group runs its payroll in this kind of tool, the agent will arrive through the vendor's update, and the three-level setting has to be decided before switching it on.

3. Reporting AI incidents becomes a shared demand. On October 9, the White House let it be known, with no published text, that AI companies must "immediately disclose" incidents involving their models, according to PYMNTS, citing Axios. On the 13th, more than 400 researchers called for an international body that records every loss of control over a model (BetaKit); in London, OpenAI and Anthropic backed binding rules, Google DeepMind and Meta declined to back new legislation (Wired). In Europe, the AI Act already requires providers of the largest models to report their serious incidents (Article 55). Why it matters: a brand that buys an agent can require its vendor, by contract, to report any incident within a set time frame.

In an interview published by WWD on October 7, Luca de Meo, Kering's CEO, defines the creativity of a Luxury brand as a chain: the studio, merchandising, purchasing, suppliers, in-store presentation. "One of my priorities in the next six months is to reengineer that," he says, aiming to open the studio up to the rest of the group.

He mentions AI only once, about the services the group wants to pool, outside the studio: "Imagine that I have to build an AI backbone for the houses. Do you want to do it 10 times, or do it once?" according to WWD.

On October 9, Saint Laurent announced the departure of Anthony Vaccarello (Kering press release, Il Sole 24 Ore). Any creative leadership that leaves takes with it part of what was never written down.

What can a shared backbone carry? Epoch AI, a nonprofit research institute that measures the progress of AI from data, offers a clue. On October 8, it evaluated models not on standardized tests, but on its own work, with its own rules. An exam tells you whether a model can do something; this test tells you whether it works the way the organization that employs it does.

Kelly Hong and Greg Burnham gave six models eleven of the institute's tasks: graphics, short analysis pieces, data tools, investigations into AI data centers, study design. Each model worked alone, as an agent, with Epoch's style guide and templates; a single human grader, one attempt per model.

What succeeds. Code and calculations are, in the most advanced models, "consistently accurate," according to Epoch.

What fails. What the authors call implicit conventions. The models pick topics too specialized for Epoch's audience. They pile on information where the institute insists on saying little, and clearly. They copy a visual code without understanding the rule behind it. The authors sum it up: the models "consistently missed implicit conventions that humans would pick up," according to their October 8 publication.

The failures Epoch found, choosing the angle of a text for its audience, saying little rather than everything, applying a visual code while knowing why it exists, threaten the tasks named at the start of this edition: translating, describing, answering. Gucci and Saint Laurent do not do them the same way, and that difference is precisely what an agent loses along the way.

What the measure does not prove. Epoch measured its own research institute, whose audience and rules are not those of a Luxury brand. The authors say so: small sample, subjective grading, scores too noisy to be taken at face value. I found no follow-up that discusses this measure with an outside view: the evidence rests on a single source.

And the line is moving: publishing an article on Substack, a task failed earlier this year, now succeeds, according to the same Epoch publication.

In Luxe Agentique, my upcoming book, I start from the five invariants of Luxury: Excellence, Exception, Experience, Emotion, Elevation. As I wrote in Edition 21, the agent can carry on its own the Excellence of execution of an office task, a calculation or a piece of code, not that of the craft gesture: that is what Epoch sees succeed. The other four require a human hand, and what Epoch sees fail belongs to them: the angle chosen for an audience, restraint, a visual code understood.

On October 6, Meta and Sierra, a maker of customer service agents, announced the Personal Agent Protocol, a common rule for the personal agent that shows up at a company on behalf of its client (Sierra, PYMNTS). The client chooses what the agent can read or do; the brand sets what it opens and what the agent is allowed to do there. On the 9th, a draft followed, with 35 new design partners, including Nordstrom, Visa, Mastercard and Adyen (Sierra, October 9, PYMNTS). It is only a draft, and no Maison appears among the 42 partners, the seven launch partners included.

On October 13, JPMorgan Chase drew its line. The bank blocks personal agents such as Meta's Muse, or Spear Street Technology's Instinct, which log in with the client's credentials to pay or transfer money; it allows them read access only (CNBC). "There's a liability transfer," Jamie Dimon warns, addressing the client who delegates (Banking Dive).

The same day, American Express published a framework in which the client records an intent before the purchase (merchant, budget, preferences); the card can decline a purchase that departs from it. Amex counts thousands of agent purchases every week, with no denominator, according to Digital Transactions.

Three players, one grid in three questions. At JPMorgan, according to Jamie Dimon, the client who delegates bears the risk; at Amex, the card blocks the deviation before the purchase; who answers if it lets it through, the release does not say. A client's mandate, like a brand's guidelines, protects only what is written.

A brand subtitles its ambassador's film, or dubs it with other voices; synthetic voice vendors promise to keep the ambassador's timbre in every language. I found no Maison that has announced it.

TED already does it. Its head of localization, Cissy Yun, described on October 7 in MultiLingual a start with eight talks in five languages, with each speaker's consent. She details one misfire: a first Mandarin version with a foreign accent, since corrected by new models.

The right ground: ambassador films and campaigns, with the ambassador's consent; in the workshop, the artisan keeps their own voice, subtitled. In May, in Edition 11, I wrote that a brand's sonic identity would become codifiable like a color. TED's experience tempers that forecast: timbre is quickly tuned, the feel of a language still needs a human ear.

First trap, consent. On October 13, according to the Asahi Shimbun, the Japanese actors' union adopted guidelines, to be published shortly: prior consent at every stage and, for multilingual dubbing that keeps the actor's timbre, additional consent and pay for each language.

Second trap, the origin of the voice: on October 9, Royal Mail acknowledged that the voice-over of its 2025 Christmas ad came from an AI; the daughter of a narrator who died in 2014 says she hears her father in it, which the brand denies, on the word of its suppliers, according to The Independent. Since August, the EU AI Act requires disclosure of a synthetic voice that imitates a real person (Article 50).

By next year's holiday campaigns, I expect a Maison, or its fragrance licensee, to have its ambassador speak several languages through a synthetic voice.

The philosopher Friedrich Schleiermacher set out in June 1813, before the Berlin Academy, the choice every translator faces: bring the reader to the author, or the author to the reader. An agent always chooses the second: it brings the brand closer to the most common language. I would take a third way: train the agent in the Maison, by writing down for it what can be written, and entrust what gets lost along the way to those who carry it. Luxe oblige !