In September 2024, Matthew Gallagher launched Medvi, a platform selling weight-loss medications, with $20,000 and a laptop. By 2025, the company was generating $401 million in revenue, with his brother as the sole employee: coding, advertising, customer service, and performance analysis were all delegated to agents…
New Mobility
Nathan Stern
The case is not without ambiguity: the FDA has issued warnings regarding some of Medvi’s advertising practices, and exploiting regulatory gray areas undoubtedly contributed to this growth. But that is precisely what makes it a revealing case study rather than a perfect model: even in a constrained and monitored environment, agentic speed proves to be formidable. What Medvi did with weight-loss medications, others will do elsewhere, whether they are better regulated or not.
Two shifts make this possible. First, the cost of collecting and processing data has collapsed, putting personalization within reach of the smallest merchant. Second, large language models now allow an agent to retain a customer’s entire history—purchases, preferences, constraints—and factor it into every recommendation, whereas a human sales assistant only remembers fragments of a conversation. Taken separately, these two shifts would be anecdotal. Together, they raise a simple question: how much of our customer relationship relied on the fact that they couldn’t know everything, compare everything, or remember everything?
Modern commerce was built on three asymmetries: the retailer knows the products better than the customer, knows the market better than them, and shapes an environment designed to guide their decision. Two mechanisms fuel this, as documented by Daniel Kahneman: impulse—the right stimulus at the right time—and least effort—stopping at the first reasonable option rather than comparing, because comparing carries a real cognitive cost.
An agent making decisions based on objective criteria does not experience moments of weakness at the checkout and does not give in to a promotion. The question is not whether these asymmetries are illegitimate—they have always been part of commerce, and a brand’s excellence also stems from its identity, its logistics, and its capacity for innovation. The question is more precise: what percentage of our revenue would not survive a perfectly informed customer? And is this share a problem to be solved, or a reality to be accepted while redeploying our efforts elsewhere?
Zero-click commerce is taking hold: the consumer states an intention, and their agent compares, verifies, and purchases without visiting a single website. During Cyber Week 2025 alone, AI agents influenced $67 billion in sales worldwide, representing 20% of global orders.
In this world, digital merchandising, visuals, and promotions no longer reach the final buyer—they reach an intermediary that can only be convinced by precise and comprehensive data. This represents a genuine loss of leverage. But for anyone who has always wanted to compete on actual quality rather than staging, it is also an opportunity: an agent making objective choices is completely indifferent to form, images, or colors. Only facts are taken into account. The question then becomes: if an agent had to judge our offering tomorrow solely on these criteria, would it be ready, or would it first need an upgrade?
Shein, with 16,000 employees, is present in over 150 countries. Temu, with 23,000 employees across 80 countries, generates $47 billion in transaction volume. These two players have shown that global scale can be achieved with tiny workforces by leveraging agents capable of exploiting purchasing behaviors at a speed no traditional human organization can match.
What they did in apparel, an agentic-native player can do elsewhere. The barrier to entry is no longer capital; it is the speed of adoption. One question remains for every brand: if such a competitor appeared in our category tomorrow, what exactly would their advantage be—and is it an advantage we could, or would want to, challenge?
According to Morgan Stanley, purchasing agents could account for $190 billion to $385 billion in online spending in the US by 2030, with food and daily essentials as the primary drivers—precisely the categories where impulse buying dominates today.
An agent that knows a customer’s true context—budget, constraints, preferences—can just as easily recommend a product as advise against a purchase. It will mechanically guide them toward retailers that demonstrate, backed by data, their actual relevance to that customer. Not out of ideology, but out of calculation. The resulting question is not about our intentions, but about our proof: if an agent had to evaluate tomorrow how well we actually serve our customers’ interests, what data would it rely on—and what would it say?
These questions do not call for theoretical answers, but for practical work: cleaning and structuring product data, customer data, and the company’s knowledge base; understanding what platforms and competitors are already preparing; identifying the first testing grounds… Several retailers have already begun to commit to this—without waiting to have all the answers, because waiting means letting others find them first. This transformation doesn’t happen overnight. But it starts, very concretely, with a conversation.
We are entering a world where the consumer’s agent and the retailer’s agent negotiate directly, without human interaction—or only at the very end, during the formal validation step. In such a context, the question is no longer just “how do we attract the customer?” but “how do we get chosen by their agent?”—and “what agent should we deploy to defend our interests in this automated negotiation?”.
This requires knowing, before an algorithm decides for us, what we are truly defending: our service commitments? Our pricing criteria? And also what we refuse to promise because we could not keep it. GEO (Generative Engine Optimization) is changing the game. Its objective: to be well-represented in the responses generated by language models, whether they are addressing a human reader or their agent. Our commitments and our criteria are no longer just topics for strategic reflection. They become the data that the algorithm will evaluate. Do we already have the answer—or are we going to improvise it under pressure?
When a consumer delegates chore-shopping—grocery shopping, cleaning products, replenishment—to an agent, they don’t lose their desire for commerce. They free up time for the commerce they choose to experience. This is the shift from forced commerce to desired commerce.
Since 2020, more than 3,000 fashion stores have closed in France, and the retail storefront footprint has shrunk by 16% in five years. At the same time, according to the WHO, one in six people worldwide suffers from persistent loneliness. These two facts do not automatically offset each other—but they frame a question: what would people look for in a physical place once they are no longer forced to go there? And do our stores today already answer this question, or only the old one?
