Two findings complicate the answer. The same customer service agent, handling returns, exchanges and order tracking, costs 100 euros a year in model fees, or 11,230, depending on the model chosen. And I have found no agent return in Luxury measured against a control group.
This edition sets out to give the executive committee what it needs to decide, in four steps: what an agent costs over three years, what it returns, how to prove it, and how to decide.
Cover: Tom Cruise in "Jerry Maguire," Cameron Crowe, 1996.
Tom Cruise plays Jerry Maguire, a sports agent who has lost every client but one. On the phone, that last client makes him shout the only line that matters to him: "Show me the money!" That is the question this edition puts to every AI agent.
Welcome to LUXE ÆTERNAI, my weekly decoding of what AI agents change, or don't, for Luxury Maisons. I am Michaël Tsakiris. I help brands put AI to work for their clients, their teams and their products, choose the right tools, and adopt them. Enjoy!
LUXE ÆTERNAI comes out every Thursday at 8:30 a.m., on LinkedIn.
LUXE ÆTERNAI comes out every Thursday at 8:30am, on LinkedIn. Subscribe on LinkedIn
The essentials of the week ¶
The thesis. An agent is judged by its return: the gain it delivers, measured against a control group, set against its cost over three years.
What I recommend. No agent in the 2027 budget without its investment case: the target return, its threshold, its metric and its measurement method; seven cost lines over three years; a verdict.
As AI gets cheaper, it gets riskier. Gartner puts an open model at 5 to 6 times cheaper per task than a frontier model, and AI strategy miscalculation enters the ranking of emerging risks at number three. I see in this the temptation to hand AI far more.
Hyatt still sells little through AI, Mistral opens its model. Hyatt still receives very few bookings through AI platforms, but judges the stakes high in Luxury; Mistral opens its large model.
Three gains announced, none verified. Amazon (50% more per order), The RealReal (20% more per transaction) and Estée Lauder Companies (no figure): none is measured against a control group.
What an agent costs and returns. What it returns: the closest gain, reported by the vendor, is worth 1 to 1.6 advisor positions. What it costs: for a customer service agent, 100 to 11,230 euros a year in model fees, or 6,200 to 62,400 euros bought off the shelf, before discounts. With editing, at 8 actions per request, the gain barely covers the purchase price.
An agent's investment case. One page, filled in by the business owner: seven cost lines, the target return, who answers when something goes wrong, three criteria and the verdict.
As AI gets cheaper, it gets riskier ¶
First fact. A Gartner note, reported on 3 November by The Register and Dataconomy, estimates that an open model costs 5 to 6 times less per task than a frontier model.
Second fact. On 2 November, the same firm published a survey of 308 risk managers, auditors and executives: AI strategy miscalculation appears for the first time, ranked third among emerging risks (CDO Magazine).
Gartner defines it as "investing in AI too aggressively, selecting unsuitable use cases or moving ahead before organizations are prepared."
Prices are falling and risk is rising: one explains the other. When AI costs almost nothing, the question of its price disappears, and brands entrust it with ever more. Each agent seems too cheap to deserve a business case. Strategy errors thus pile up, quietly, and they are what will prove costly.
I draw one rule from this for the next budget: an agent's expected return is written down before it launches, even when it costs little.
Sources for this section4 links
Hyatt still sells little through AI, Mistral opens its model ¶
Hyatt: few bookings, high stakes
On 29 October, Mark Hoplamazian, Hyatt's chief executive, described bookings made from AI platforms as "very, very small." But, he added, "it matters a lot in luxury and lifestyle" (Earnings Whispers, The Motley Fool, Yahoo Finance).
For the next media plan, ask your agency for the share of sales that came through an agent, channel by channel.
Mistral opens its large model
On 29 October, Mistral published the weights of Mistral Large 4: a company can now run it on its own servers (Mistral, TechCrunch, Let's Data Science).
Three gains announced, none verified ¶
Three companies announced AI-related gains this week. None measured them against a control group: nothing yet shows what an agent returns.
Amazon: self-selected customers
On 29 October, Andy Jassy announced that more than 400 million customers had used Alexa for Shopping, Amazon's shopping agent, and that its US users spend 50% more per order than other customers, according to the call transcript and Modern Retail.
The figure compares customers who chose the agent with everyone else. Nothing shows that the former were not already spending more.
The RealReal: the early adopters
The Luxury resale platform opened Ask TRR, its search agent, to its members in October (Edition 26). In its third-quarter results, it writes that adoption "is still building" and that "our early adopters spend 20% more per transaction," according to its prepared remarks and the transcript published by TradingKey.
Same kind of evidence: early volunteers compared with everyone else.
Estée Lauder: no figure published
On 30 October, the group announced a long-term partnership with Microsoft: Copilot is being extended to more than 14,000 office employees, according to the group's press release, Microsoft's article and Chain Store Age.
No amount, no usage, no return.
What I recommend
- At every AI license renewal, ask the vendor for the number of weekly active users, tool by tool, and negotiate seat prices on that basis.
- For every return figure you are shown, ask who measured it, and against which control group.
What an agent costs and returns ¶
1. Seven cost lines over three years
An agent costs far more than what it takes to run. Over three years, its cost follows its life cycle, in seven lines:
| Line | What it contains | Expensed or capitalized |
|---|---|---|
| 1. Scoping and design, one-off | Interviews and on-the-job observation, use case sheet, state of the data, measurement set before launch, rules and safeguards | Usually expensed |
| 2. Tuning, integration and testing, one-off | Tuning on documents and data, connection to orders and inventory, test set built on real tasks | Capitalized if the agent is built and the company controls its code; expensed if it is bought |
| 3. Adoption and training | Team onboarding, usage rules (who checks, who signs off, who answers), team training | Expensed |
| 4. Running, every year | Model or billed actions, hosting, advisors' review time, security | Expensed, except servers bought outright |
| 5. Upkeep, every year | Test set rerun at every model change, fixes, monitoring of users and of review costs | Expensed |
| 6. Legal and compliance | Impact assessment, contracts with each provider, customer disclosure, registration if the use is high-risk | Expensed |
| 7. Exit, one-off | Data to retrieve, access to shut down, team to rebuild | Expensed |
Only running costs have public prices: those of models and of agents bought on a usage basis. The model price comes as a range, set by three variables:
- volume: how many tasks per day;
- the model: its price per million tokens, the token being the word fragment on which an AI model is billed;
- the billing method: per token, per seat, per action, per conversation or per outcome.
The case: a brand's customer service agent, where 20 advisors answer a few hundred requests a day. It takes the 100 simplest, a return, an exchange or order tracking:
- it reads the order, applies the return policy, checks inventory, carries out the operation under a refund cap set by the brand and drafts the reply;
- the best clients, recognized from their profile, and sensitive cases stay with the advisor.
Calculation: 8 model calls per request, 2,000 tokens read and 500 written per call, 350 days a year. Vendors' list prices (for example Google), with no caching or discount, converted to euros at the European Central Bank reference rate of 2 November, 1.1216 dollars to the euro (ECB).
| Model (vendor) | Price per million tokens, read / written | Per year |
|---|---|---|
| Gemini 2.5 Flash-Lite (Google) | €0.09 / €0.36 | €100 |
| DeepSeek V4.1 Flash (DeepSeek), peak hours | €0.27 / €1.07 | €300 |
| Mistral Large 4 (Mistral) | €0.62 / €1.87 | €610 |
| Sonnet 5.5 (Anthropic); Gemini 4 Argon until 30 November (Google) | €1.78 / €8.92 | €2,250 |
| Opus 5.5 (Anthropic); Gemini 4 Argon from 1 December (Google) | €3.57 / €17.83 | €4,490 |
| GPT-6 Astra (OpenAI) | €8.92 / €44.58 | €11,230 |
Prices checked on 3 November: OpenAI · Anthropic · DeepSeek; Mistral: announcement cited above
The model remains the smallest line. Design, integration, testing, training and the way the agent is bought weigh more.
Being able to switch providers
Between the agent and the models there is often a platform: LangChain to assemble and monitor the agent (35 euros per user per month), Dust to hand it to teams (30 euros per user per month), Hugging Face to host an open model.
These layers cost little, and they decide the dependence. A platform that routes each step to the chosen model lets you switch providers the day its price doubles. That is economic sovereignty: being able to leave.
Tune, buy or build
Three paths lead to this agent.
- Tune the assistant you already pay for: the cheapest option, limited by what the tool can do.
- Buy the off-the-shelf agent from a major vendor. Salesforce's prices for Agentforce come to, for the 35,000 annual requests, 6,200 to 25,000 euros a year on per-action billing (0.09 euro, the price noted in Edition 28, 2 to 8 actions per request), and 62,400 euros on per-conversation billing (1.78 euros according to its price list).
- Have it built: the model weighs least here, upkeep most, because every new version of the model or the platform forces the control tests to be rerun.
Two European regulations to budget for
- The General Data Protection Regulation (GDPR) may require an impact assessment when an agent profiles customers at scale, and a contract with each provider that touches their data; a model hosted outside Europe adds a transfer framework, and China has no adequacy decision.
- The AI Act already requires an agent's provider to inform customers that they are talking to an AI; the brand that buys it has every interest in writing this into its contract (Morgan Lewis); the obligations for high-risk systems apply from 2 December 2027 (Future of Privacy Forum, Legal500).
Training a model: rarely
A company that deploys an agent does not train a model; it can fine-tune one on its own examples, for 1.34 euros per million training tokens at Google, and above all in internal time: annotating, testing, readjusting at every model change. For the customer service agent, written rules are enough: the line stays at zero.
Bought, the agent weighs on EBITDA
Under IFRS (International Financial Reporting Standards), the cost of configuring a cloud-based service is generally expensed, because the customer does not control the provider's tool (IFRS Foundation, Grant Thornton).
At equal total cost, a bought agent hits that year's EBITDA almost in full; a built agent is partly amortized.
The cost is therefore a range.
2. What it returns, on three levels
Three levels, from the easiest to measure to the most useful for the budget.
The first level steers the agent: the cost per completed task and real adoption (weekly active users, relative to paid seats).
The second level measures business value. Four metrics apply in every area, measured before and after:
| Business value | What is measured | The trap |
|---|---|---|
| Quality | The share of the agent's work accepted without edits by the business expert | The vendor's satisfaction score |
| Time saved | The hours given back to the client, the craft or the trade, clocked | The automation rate |
| Fewer errors | Errors per thousand tasks, logged in a register | The volume of tasks processed |
| Lower risk | Compliance or security incidents, and the time taken to install patches | "The agent is reliable," with no incidents counted |
Each of a Maison's seven areas adds its own metric, measured with a control group, and a mirage to avoid:
| Area | The value metric | The mirage |
|---|---|---|
| Customer experience and commerce | Assisted conversion: sales the agent prepared and the advisor closed | The number of queries handled |
| Brand, narrative and sovereignty | Fidelity of the narrative agents relay: the brand told in its own words | The raw number of citations |
| Communication and visibility | Qualitative share of voice, coverage earned without being paid for | The volume of content published |
| Employees | Retention and engagement of the teams the agent assists | The number of copilots switched on |
| Operations and supply chain | Stockouts and markdowns avoided | The volume of automated tasks |
| Creation and product | Share of the agent's proposals taken up by the designer | The number of images generated |
| Talent and transmission | Retention of new hires, internal mobility revealed | The number of résumés screened |
| All areas | The area's metric, above | The number of tokens spent |
The third level translates business value into money; it decides the budget. The gain is measured in additional margin, against the control group:
- time saved counts if it is redeployed to selling or saved in paid hours;
- an error avoided is worth the cost of correcting it;
- a reduced risk is recorded separately, in incidents avoided, without being added to the gain.
The ROI (return on investment) over three years is then calculated as follows: margin gain, minus the total cost of the seven lines, divided by that total cost. It comes with the payback period, the number of months before cumulative margin covers cumulative cost, and is recalculated in the high scenario.
The break-even point
Let's apply the grid to our agent. The quantified gain closest to this case comes from Salesforce (recalled from Edition 28), for L'Oréal's consumer care service in North America, where the agent drafts and the advisor reviews: the handling time of a case drops from about 7 to about 2.5 minutes without edits, and by about 40% with edits (Salesforce).
- The gain. Applied to the 35,000 requests, this benchmark frees up 1,633 to 2,625 hours, or 1 to 1.6 positions at 1,607 hours worked a year (Service-public).
- Its value. At the salary of a customer service advisor at a Luxury Maison in Paris (29,000 euros gross a year, Sintel Recrutement posting, September 2026), that is worth 29,000 to 47,000 euros a year, plus employer social charges, provided those hours stop being paid.
- The cost. The bought agent at 8 actions costs 25,000 euros a year.
- The ceiling. Over three years, 13,000 to 67,000 euros remain for design, integration, testing, training, upkeep and legal: that is their ceiling for a positive ROI, wider at 2 actions per request.
If every draft is edited, the gain barely exceeds the purchase price: editing and the number of actions decide, far more than the salary.
Built, the agent costs 1,830 euros in model fees over three years with Mistral Large 4, 6,750 with Sonnet 5.5: its construction and testing set the threshold.
This threshold is written down before the pilot; time per request, measured against simple requests drawn at random and left to the advisors, shows whether it is crossed.
Above everything, a veto: the full-price sell-through rate, tracked by collection, every month. A positive ROI achieved by degrading it does not count: the agent has gained margin by damaging the brand.
3. Four levels of evidence
This gain is reported by the vendor. Yet a gain is only as good as its evidence, which four levels rank:
| Level | Who measures, and how |
|---|---|
| 0 | The announcement, with no published measurement |
| 1 | The user company, or a survey, with no control group |
| 2 | The company, with a control group and a baseline period |
| 3 | An independent third party, on a described sample |
In Luxury, the McKinsey report on agents publishes no measured return.
Without a control group, a return remains a hypothesis: it opens a capped pilot budget, released in tranches at milestones, with the gain entered at zero.
4. Four possible decisions
Cost and evidence combine into four decisions, depending on two questions. Is the cost quantified over three years, from the low scenario to the high? Is the gain measured against a control group (level 2 or 3)?
| Gain measured with a control group | Gain announced, not measured | |
|---|---|---|
| Cost quantified over three years | Scale up: business owner, real data, three-year budget | Extend the pilot by six months, with a control group and a written threshold |
| Cost still unknown | Quantify the cost, then scale up or slim down: model, volume, billing | Pull the plug, or don't start |
What I recommend
For our agent: first quantify its six other cost lines, then pilot it for six months against a control group.
An agent's investment case ¶
Like any investment, each agent has its case: one page, filled in by the business owner, reviewed by finance and human resources before the budget is approved.
- The agent. The task it performs, in one sentence; the business owner accountable for it, named.
- The seven lines, over three years, as ranges, each marked as expensed or capitalized after review by the statutory auditors.
- Two scenarios. The high one: the model price doubled, as Gemini 4 Argon's will be on 1 December (usagepricing, ActuIA). The low one: the maximum loss if the agent is stopped at six months.
- Steering and business value: the six common metrics described above, tracked every month.
- The target return. The value metric of its area, its measurement method with a control group, its level of evidence today, and the three-year ROI with its payback period.
- Three criteria, scored from 1 to 4. Reversibility: the agent can be unplugged and the data comes back. Dependence: the number of vendors whose pricing can change the total. Accountability, below.
- Who answers when something goes wrong. An agent acting on the Maison's behalf commits it to the client: an undue refund, disclosed data, a wrong promise. In 2024, a Canadian tribunal held Air Canada liable for false information given by its online assistant, rejecting the idea that the assistant was a separate entity (American Bar Association). The case names the owner of each risk, what the vendor's contract covers, and the insurance.
- The verdict, one of the four decisions, and three dated milestones: at six months, usage reaches the threshold written before the pilot, or the agent is stopped; at twelve, a first return measured against a control group; at eighteen, revise or expand.
In the Harvard Business Review, Professor Das Narayandas describes a bank where, six months after a credit decision system was approved, most files still went through human review. "Authority fails to move when people remain personally liable for the system's outcomes," he writes.
Intended verification is a chosen cost; an agent's work redone out of distrust is a cost suffered, the same decision paid for twice.
The case carries Narayandas's four questions, with an answer to each:
- who loses a share of decision-making if the system works;
- who gains one;
- who explains the outcome when the system was followed;
- who carries the risk during the learning period.
If a line stays blank, the agent is not ready for the budget.
Sources for this section4 links
Three open-access reads ¶
- The study. "Open-weight models and the patch gap," Epoch AI, 29 October. An agent built on an open model turned Firefox patches not yet shipped into attacks, for about 900 euros per version.
- The transcript. Marriott's third-quarter analyst call, The Motley Fool, 2 November. AI is changing how travelers discover a trip before they book it.
- The press release. Zalando's third-quarter results and the launch of Zalando Extended, EQS News, 3 November. An agent that will shop at other merchants.
Tapestry reports on 5 November ¶
-
Thursday 5 November, 8 a.m. in New York.
Results for the first quarter of fiscal 2027 of Tapestry, owner of Coach and Kate Spade, according to Finviz. To listen for: a return figure on agents, and its level of evidence.
A Luxury Maison costs out a bag down to the last cent, and promises in its window only what it has seen leave the workshop.
It owes the same rigor to every agent, wherever the agent must serve:
- its clients, measured by full-price sales;
- its teams, measured by weekly usage;
- its products, measured by the share of the agent's proposals the designer takes up.
Choosing the right tool means quantifying its cost and keeping the freedom to leave. Adopting it means measuring its return before promising it. Luxe oblige!
Sources for this section1 link
LUXE ÆTERNAI newsletter
Get the next edition
Every Thursday at 8:30am, what AI agents change, or don't, for luxury Maisons. You subscribe on LinkedIn, in one click.
Michaël Tsakiris
Paris, Thursday