The system prepares. A person decides.

There is a pattern in the examples of AI that work in luxury, and it is not the one the vendor decks suggest.

Tiffany has described advisers working in an AI-assisted platform: a chat interface that holds client preferences, product information and store procedures, with the adviser continuing the relationship. Louis Vuitton, working with Comelz, has shown hides being digitised, their characteristics identified and cuts proposed, for the artisan to validate or adjust. In one of the largest field studies of generative assistance at work, 5,172 customer-support agents resolved about 15 per cent more issues an hour, with the largest gains among the least experienced and small quality declines among the most skilled. The agent still answered the customer.

Three examples, one of them outside luxury, and the same pattern in each: the system prepares and a person decides. That is not a limitation waiting to be engineered away. It is the design, and when a proposal breaks it, that is the first thing to notice.

I knew this before I had read any of it. [Knitting floor: the programme says one thing; the technician who has run that yarn on that machine says another; the garment is made the second way, and the record should show why. Replace with a true instance from your own factory, two or three sentences.]

What "a person decides" has to mean

It is easy to put "human in the loop" on a slide and harder to say what the human is allowed to do. At least four things.

The adviser can decline. A suggested message to a returning client is a suggestion. The adviser inspects the facts behind it, judges timing and occasion, and may reject it, rewrite it, or decide not to make contact at all. If the workflow measures advisers on messages sent, the right to decline exists on paper only.

The artisan validates the cut. A proposed cutting layout is a proposal. The person who knows what that hide will do under a needle has the last word, and the system records the override rather than penalising it.

Creative direction stays human. Generative tools widen exploration and speed iteration. They also converge: in a randomised experiment with 293 writers, access to generated ideas raised individual creativity scores and made the stories more alike. Eight-sentence fiction is not fashion design, but the tension is real. A house that uses these tools needs its own review for distinctiveness, run by people, against its own archive.

A service owner can stop it. Somebody with authority can suspend a feature that is producing wrong answers or unwanted contact, without a change request to the vendor. If nobody holds that switch, nobody owns the service.

If any of the four is missing, the design has a problem, however good the model.

A worked design

Suppose a house has piloted an AI clienteling tool and the vendor is pressing for rollout. Here is a design that would satisfy the pattern. It is a proposal, not a description of anyone's system.

The system prepares. For a returning client, approved product facts and permitted preferences are retrieved; search does that, and search does not need a model that learns. Predictive AI ranks candidate products or moments. Generative AI drafts a message, if it is used at all. Each function is justified separately, because each fails differently, and each depends on the product record being right, which is the previous note's subject.

The adviser decides, between three outcomes: a helpful contact or appointment, prepared and welcome; no contact, because circumstances have changed or the moment is wrong; or escalation to a specialist, because the client has a care question that needs an expert. Training time is allocated for this. Product specialists correct information that is wrong.

What you measure: accuracy of the prepared facts; voluntary adoption by advisers at six months; perceived attention from clients; unwanted-contact complaints; repeat behaviour. All of it against existing service without the tool, and against the simpler alternative of better information or training, after staff time and operating cost.

More messages, sent faster, would not settle the question, and more messages is what a system optimised for messages will produce.

The limits the design has to carry

Luxury makes four limits expensive, and the design above has to answer each. Fit: a render cannot prove fit, drape or feel, so a viewer should say what it represents and send the client to a hand or an adviser. Truth: a generated answer can invent an attribute, which is why the facts are retrieved from an approved record and corrected by a specialist. Privacy: recognition must never feel like surveillance, and since August 2026 the EU requires people to be told when they are dealing with an AI unless it is obvious. Authorship: prompting alone is not authorship, which is why creative direction and the review for distinctiveness stay with people. A better model solves none of these. Deciding who holds the decision does.

The question to ask

When the next proposal arrives, do not start with the model. Start with the person. Who prepares, who decides, what can they decline, and who can switch it off. If the answer to any of those is "the system", what you are looking at is not assistance but a replacement that nobody has called one, and your clients will judge it as a replacement.

---

---

Sources