No Such Thing as a Free Agent

• The thesis. Deploying an agent that applies a brand rule means engraving that rule into the machine and running it without you. The question of the year is no longer the tool or visibility: it is governance, who writes and who answers for the rule the machine executes. The concept: the compiled brand, doctrine become code.
• The paradox. We hand the machine the enforcement of a brand policy we never wrote. The agent fills the void with a default doctrine no one has signed.
• What's moving. In luxury hospitality, an agent now sells a stay end to end, from request to booking. At Louis Vuitton, the doctrine fits in three words: "augmentation, not automation."
• Beyond Luxury. McDonald's is testing a voice agent at the drive-through (five restaurants; about 90% of orders with no human stepping in across more than a million handled, figures from a franchisee, unconfirmed by the company); an administration filters access to the best agentic engine; and for the first time, bots outnumber humans on the web, 57.5% of traffic.
• The real battleground: cost. The price of running an agent is collapsing in a price war between OpenAI, Google and Anthropic; the same month, we learn that access to these engines can be cut overnight. Uber burned through its entire 2026 AI budget in four months.
• What an agent will not do: write the rule. It applies it, and fills it in when it's missing. The doctrine remains for a human to sign. The Agentic Delegation Plan, below, is the method for writing it before the agent does.

We have handed an agent the power to enforce a brand law that no human ever voted.

The scenario repeats, Maison after Maison. You deploy a clienteling agent, a compliance agent, a sorting agent; you set it, plug it in, let it run. The question "which rule, exactly, is this agent applying, and who wrote it?" is never asked. Yet the rule exists: there has to be one for the agent to return a verdict. It's just that the Maison didn't write it. The software vendor did, in its default settings; the language model did, in its reading of what "luxury" is; the engineer did, by fixing a threshold no one was asked to approve. Three authors no one at your Maison has ever met write your rulebook together: the grid of what passes, what's refused, and at what threshold.

This is the paradox of the absent legislator: we delegated the enforcement of the law without voting the law. And because the agent reasons, it does not merely execute: where the text is missing, it invents the missing clause, silently, recommendation after recommendation. The brand ends up written. By someone. The question is whether it's you.

Three dated moves this week, read through the same lens: who writes the rule the agent executes, and who answers for it.

In luxury hospitality, an agent sells a stay end to end. Unveiled in mid-June at HITEC, the global hospitality-technology show, an Agentic Sales Coordinator captures a group or event request, qualifies it, follows up and carries it through to a confirmed booking, with no human intervention. Its maker, Canary Technologies, already equips more than 20,000 hotels and counts Marriott and Wyndham among its partners. Selling a suite or privatizing a floor means applying a thousand tacit rules: who you call back first, what you grant, in what tone you decline. An agent that closes the sale applies those rules too, except that today no one dictated them to it: it reconstructs them from a booking history, not from the Maison's doctrine of hospitality.

Another hospitality player shows where the agent must stop. At the same show, the vendor Trybe presented an agent that sells a spa package and dispatches it on its own across the spa, the hotel system and the restaurant; in a real British case, the Woolacombe Bay hotel saw spa bookings rise 14% in three months after adding a forecasting AI. The instructive part is the limit the vendor sets itself: it refuses to hand the agent pricing, judged too sensitive to run without a human. There is a rule written where it counts: not "the agent can do anything," but "the agent goes this far, and no further." The boundary is set by decision, not by default.

LVMH makes the agentic a doctrine of sales, with a reservation. In June, the group's chief information officer called agentic AI "the future of selling," describing a layer of agents that converse with client data and open up leads the advisor would not have seen. But the rule, at LVMH, is written and held: "it's mainly about anticipating, understanding, interpreting, responding to our client's intent… powered by AI," sums up Louis Vuitton's global head of client development, Soumia Hadjali; "it's more about augmentation, not automation." The agent equips the advisor, it does not replace them. That is precisely the discipline missing elsewhere: a rule set before deployment, not discovered after.

Three agentic facts of the week, beyond Luxury, that you won't find anywhere else in this edition.

1. McDonald's is testing drive-through ordering by an agent. Around its worldwide convention, in June, the chain unveiled ArchIQ, its AI platform built on Google's cloud. On the customer side, a voice named "Archy" takes the order at the drive-through, answers in English and Spanish, recognizes a regular and their usual order, and works alongside the crews rather than replacing them. The test covers only five US restaurants, but the figures put forward, by a franchisee and not by the company, set the tone: more than a million orders handled, about 90% with no human stepping back in. What matters for a Maison: even on a pilot, the operating rule is encoded once, then applied identically. The compiled brand is not waiting for 2027; and the rule the agent executes, someone wrote it, McDonald's or Google.

2. An administration now chooses who gets access to the best agentic engine. In late June, OpenAI reserved access to its most advanced model, named GPT-5.6 Sol (the one that leans on sub-agents for heavy tasks), to a couple dozen partners approved by the US government. The stack on which your Maison's agent will run tomorrow, the pile of models and software that makes it work, stops being a mere technical choice: it becomes a dependency a foreign state can open or close. The sovereignty of that stack, which I treated in #17 as a theoretical risk, has just become a dated fact.

3. For the first time, bots outnumber humans on the web. In early June, Cloudflare measured that 57.5% of requests to the sites it protects come from automated systems, against 42.5% from humans, and its network already returns more than a billion toll requests a day (the HTTP 402 code, "payment required") to the bots that scrape content. Your brand's first reader is now a machine, one that reads your site on behalf of a client. The unease is not new: back in 2025, LVMH's head of technology worried about it, "I'm not sure it's a good idea to have assistants systematically crawling our web platforms to buy luxury products." Deciding which of these agents you let in, block or charge becomes an act of sovereignty over your own brand data.

Most Maisons think they're deploying a tool when they are promulgating a law. Decoding the compiled brand means seeing where the line runs between the rule you wrote and the one the machine writes in your place.

Code is law. Compiling your brand means tracing the Simondon Line in code, that boundary, set in #7, between what you delegate and what you keep. As long as it stayed an idea, you could leave it blurry; compiled, it becomes an executable decision, taken thousands of times a day. The legal scholar Lawrence Lessig named this shift twenty-five years ago in a phrase that stuck: "code is law." His thesis is not that software replaces the law, but that it constrains conduct as surely as a rule does; and that, like a law, it is a choice someone is responsible for. A compiled brand doctrine lands exactly there: what the software allows becomes the rule, with no appeal and no exception. And a rule executed with no author is not neutral: it leans. The model that drives the agent has its own idea of what "luxury" is, learned on a web where the most documented brand weighs more than the most desirable one; the vendor's factory setting has its own appetite for risk. Neither consulted your committee, and yet, between them, they write your rule.

A compiled rule costs almost nothing. Once compiled, a rule is no longer a rare object: it applies at near-zero cost, as many times as you like. The economist Paul Romer, Nobel laureate in 2018, built his theory of growth on this property: an idea (a plan, a formula) is non-rival, using it does not consume it, you can replicate it endlessly. A brand doctrine, as long as it lives in an artistic director's head, is rare and rival; compiled, it becomes that non-rival good, applied in one stroke to every window, every email, every piece of advice. The value, then, no longer sits in the agent (Microsoft, Google or Amazon platforms plug in within days) but in the quality of the text you give it to execute. At that scale, an error in the rule does not cost once: it produces a drift, systematic and silent, until the day it becomes visible. And in a Maison, that drift reads not in incidents but in margin points: a badly written discount rule chips away at full price, a wrong tone in a reply erodes desirability, and those losses show up only on the balance sheet. Writing the rule costs a few committee days; living with it costs a position. The least spectacular work has become the most strategic.

But that "almost nothing" is a battleground. An agent's cost is paid in tokens, the units of text a model bills on each request, and that price has become the heart of the matter. Sam Altman, OpenAI's chief, admitted as much at a company event in early June: the cost of tokens, never raised at the start of the year, has become "a huge issue"; "the meme I see now," he said, "is: my company spent my entire 2026 budget in Q1, can you make this more efficient?" He is not exaggerating: Uber burned through its entire 2026 AI budget in four months, Fortune reports. Hence a price war opened ahead of the upcoming listings: Google has just cut its cheapest plan from eight to five dollars a month. Enough to compile your brand on the cheap; except the ground gives way. On 12 June, on Washington's order, a major vendor had to cut off two of its most powerful models for all its clients, everywhere, with no notice (I told the story in #17); in late June, another vendor's best engine is reachable only by the partners a state approves. There is the economic truth of the compiled brand: you write the rule, but you do not own the infrastructure that runs it; its price can collapse, its access can close, on a decision that is not yours. Hence the caution already set in #17: keep several exits, never hand over the only spare key.

What can't be written can't be compiled. There remains a boundary the compiling does not cross. The philosopher Gilbert Ryle distinguished, back in 1949, two kinds of knowledge: knowing-that, a stock of true statements you can write down, and knowing-how, that intelligence shown in the gesture without reducing to a list of rules. A Maison's taste is knowing-how in Ryle's sense: an artistic director's eye, an artisan's hand, a seasoned salesperson's tact. You can transcribe part of it into knowing-that (the mandatory disclaimer, the reorder threshold, the charter of a tone), and an agent will apply it better than a human, fast, tireless, with no exception forgotten. But the rest resists writing, and it is that uncodable remainder that separates a living brand from a compliance engine. An agent that "applies" a judgment you never formalized does not apply it: it guesses it, and so betrays it.

Who signed the rule? Hence the real question, which is not technical but political. McKinsey puts it bluntly: "the strategic question is no longer whether agents will mediate the journey, but who controls how brands are perceived." In a Maison, a brand decision always commits someone: a director signs a refusal, a head answers for a disclaimer. When that decision moves into an agent, the signature vanishes unless someone affixed it upstream. And the signature stops being a moral comfort: the EU AI Act now places legal liability on whoever deploys an agent, up to 7% of global turnover for a prohibited use, and 2 August 2026 marks the full application of the high-risk regime. The rulebook no one signed becomes a quantified risk, pinned to a name. A few Maisons saw it coming: Tiffany, the LVMH group's jeweler, has just added to its org chart a role dedicated to operational AI, giving a name, and therefore a signatory, to the doctrine its agents apply. That is the exception. And we already speak of machines that decide for real: at Uber, about 10% of shipped code is now written by agents on their own. Taking back the reins does not mean unplugging the machine. It means writing the law where it can be written (a choice, not a fate, Lessig would say), refusing to code it where it cannot, and putting a name at the bottom of the text. Keep your hand on what can be coded, and your pen on what cannot.

No brand has harder rules to keep than a wine and spirits group. Diageo cannot, by law and by its own charter (the Diageo Marketing Code, the internal document that sets what one of its alcohol brands is allowed to show), address its communication to anyone under the legal drinking age. "We can't market, and wouldn't want to market to anybody under the legal drinking age," summed up Joshua Nafman, its data and operations director: "so that means who we can communicate to and how we can do that, we're really challenged on that." There is a rule, written, signed, enforceable, that lived forever in a document applied by hand, visual by visual.

That rule, Diageo is moving to the other side: from text you read to a program that runs. With its ACE system (for Automated Consumer Experience), the group says it automates more than 70% of its retail creative tasks and applies brand compliance in real time, at production rather than as an after-the-fact correction, for 3.5 times more campaign launches at constant effort. These are figures communicated by the group and its vendor, which I cite under that reservation. But the fact they describe is the right one: a rule that lived in the heads of a few keepers of the Code now applies mechanically, to each piece, ahead of the human.

What makes the leap safe comes down to one word: priority. The rule existed, written and signed, before the agent. Banning any approach to minors, requiring a disclaimer, bounding what an alcohol brand may show: so many explicit, verifiable clauses that humans debated. That is exactly what an agent applies well. The Maison that lacks this text and still deploys a compliance agent does not remove the rule: it delegates it to the machine, which fabricates it as it goes, with no signatory and no archive. The difference shows on the day of a dispute, when you have to produce the rule that justified a refusal: one shows it, the other discovers that no one ever wrote it.

Compliance is probably the first ground where compiling your brand is not only safe but desirable: a domain of explicit rules, exactly what the machine executes without fail, and where it gives teams back real hours each week on repetitive tasks. The danger begins when you extend the reflex to grounds where the rule is not written: the tone of a sensitive reply, the handling of an exception, the "no" that protects prestige. There, there is no code to execute but a judgment to make; it is the knowing-how Ryle spoke of, which no charter exhausts. Technology compiles the rule; it does not promulgate it. The Maison that grasps this in time will hold the pen; the others will sign a text someone else has written.

To a Maison's leader, or to its executive committee.

Which rule are your agents applying, right now, that you never wrote down in black and white? If I asked to see the exact rulebook one of your agents executes (what it allows, what it refuses, at what threshold), could you produce it, signed and dated, or would you discover that no one ever wrote it?

The question is anything but abstract. Gartner measures that 80% of applications shipped or updated in early 2026 embed at least one AI agent, against 33% two years earlier; but barely a third have one truly in production, and nearly nine pilots in ten never get there. In other words: we plug in fast, we govern badly. The rulebook, meanwhile, built itself, setting after setting, inherited from a vendor's defaults and from a model's reading of "luxury." Writing it down, task by task, deciding for each whether it stays human-only, AI-supervised or autonomous, and settling it before the agent settles it in your place: that is the work I formalize in an Agentic Delegation Plan. Five steps:

1. Framing and prioritization. We gather your inputs (workflows, tools, data) and draw from them a prioritized backlog of tasks that are candidates for delegation. You see, in black and white, everything an agent could touch.
2. Arbitration. At the heart of the method, a delegation matrix decides each task on two axes: what pushes toward delegating, what keeps the human in. Each criterion is weighted to your Maison's needs and will: you set the cursor. Verdict per task: human-only, AI-supervised, or AI-autonomous. This is where your Simondon Line is drawn, bespoke and by hand, not by default.
3. Specifications and architecture. For each task kept, we write the real rule and the choice of tools, with your chief technology officer (CTO), your IT director and the heads of the business lines involved, as close as possible to your existing systems.
4. Implementation. The plan: the order of deployment, who keeps their hand on what, the guardrails set. You leave with the map and a sheet per task.
5. Deployment. In waves, under control, then the support takes over for as long as it takes the agents to hold at scale.

At the end, neither a charter, nor a toolbox, nor one more method: a clear, specified and prioritized action plan to implement, govern and evolve your agentic workflows. Its aim is not to make the machine do more, but to entrust it the codifiable part so as to give your teams back to Excellence: the gesture, the judgment, the relationship, what precisely does not compile. We identify the right workflows, and I help you put them in place.

Discover the Agentic Delegation Plan →

The tech making the buzz this week is a body, not software. And no one, in Luxury, expects it where it can serve.

Embodied AI, humanoid robots driven by models that see, understand and act, has just changed floors. On 24 June, Agility Robotics became the first humanoid maker listed in the United States, valued at about 2.5 billion dollars; 1X opened pre-orders for its home robot NEO (20,000 dollars, or 499 dollars a month); Google DeepMind runs its Gemini Robotics models directly on the machine, without a network. And next week, on 7 July, Paris hosts at STATION F the MACHINA Summit, the first European summit dedicated to this "physical AI." The figure that speaks to Luxury is the hand's: a human hand has 20 to 27 degrees of freedom; the NEO robot's aligns 22, with a touch able to feel a texture and adjust its grip. The robotic hand of 2026 does what the 2024 one could not.

The reflex would be to picture the robot-salesperson or the robot-courier. That's flat, and it misses the point. The new ground, for a Maison, is the back of the atelier, where the gesture is painful, repetitive or risky for the piece. Three uses, at eighteen months. The gesture repeater: trained by teleoperation on the wrist of a Meilleur Ouvrier de France (a state-certified master craftsman), a humanoid is not meant to stitch the bag, but to record the exact pressure of the saddle stitch, the angle of the awl, the tension of the thread, and to become the training partner of apprentices when the master is gone. The master's hand becomes transmissible without being replaced. The archive handler: in a heritage reserve, taking out, turning over, photographing and putting back a 1947 dress or a set piece demands a delicacy a human loses by the hundredth piece; a fine touch keeps it constant, traceable, with no snag. The window pre-setter: the night before a launch, the robot sets the heavy and repetitive elements to plan, and the visual merchandiser keeps only the gesture of the eye.

It's the rule of the inside and the outside I set in #12: the machine works the inside (the reserve, the training, the logistics of the beautiful), never the outside the client touches and pays for. The trap already has a name: teleoperation, that "autonomous" robot in fact driven by an invisible operator. The day a saddle stitch sold as "handmade" comes off a robotic arm, it's no longer Luxury, it's counterfeiting the gesture.

By the end of 2027, bet that a great Maison will deploy a humanoid in its reserve or its training atelier (never on the line of the piece it sells), and make it an argument of craft transmission, not productivity. The first to claim it publicly will turn an image risk into proof of heritage; the others will do it quietly. The machine can learn the artisan's gesture. It will never inherit the hand, and it's the hand, not the gesture, the client pays for.

• A Framework for Governing Agentic AI, Deloitte Insights, June. The missing piece of the debate: who answers for an agent when it acts alone, how high responsibility climbs, how to document and audit the rule it executes. Read it to move from "you need a signatory" to "here is how you organize it."
• Why LLMs Misunderstand Luxury Brands, Harvard Business Review × INSEAD, 22 June. The demonstration that models read the explicit perfectly (price, label, name) and stay blind to the implicit: rarity, heritage, restraint. In other words, the part of your brand no agent will apply if you don't write it yourself. The exact complement to this edition.
• OpenAI Considers Drastic Price Cuts, Anticipating War for Users With Anthropic, Wall Street Journal (via Forbes), 11 June. The piece that puts numbers on the agents' price war. Read it to understand why the economics of the compiled brand are as unstable as they are promising, and why cost, not capability, will be the arbiter of 2026.
• Humanoid Robots: Crossing the Chasm, McKinsey, 2026. Why embodied AI moves from concept to industrial deployment, and what that changes for a Maison's physical operations. Read it alongside the Beta section: the machine's hand is coming, the question is which post you welcome it to.

Next week, Paris puts the two ends of this letter on the same agenda.

• Monday 6 to Thursday 9 July, Haute Couture. The Federation opens the Fall-Winter Couture in Paris: about thirty shows, Schiaparelli and Dior from Monday, Matthieu Blazy's second couture for Chanel on Tuesday. The absolute summit of the human gesture and of rarity, exactly the value no agent will apply in your place.
• Tuesday 7 to Thursday 9 July, RAISE Week. A few streets from the shows, Europe's largest AI summit (about 9,000 executives at the Carrousel du Louvre) makes the agentic one of its three pillars; in the opener, on the 7th, the MACHINA Summit puts the spotlight on embodied AI. The hand that sews and the machine that wants to learn it share the same city, the same days.
• Thursday 9 July, the law weighs in. The European fashion and luxury law conference (IBA) puts "AI and intellectual property" on its program: who answers, legally, when an agent commits the brand. The legal frame of the compiled brand starts to be written there.

This year's Luxury plays out in the gap between those two agendas. Compile what can be compiled (compliance, replenishment, logistics) to give your teams back to what does not compile: taste, the gesture, the "no" that protects prestige. Plugging in an agent amounts to putting a law into force; the only question is who wrote it. Taking back the reins is writing the rule before you compile it, and keeping for yourself the hand that does the rest. No one is asking you to unplug the machine. Luxe oblige.