Philippe Corrot: Marketplace Model Beats the SaaSpocalypse

Key takeaways

1. The marketplace mechanism, and the figure that served as the founding thesis. The merchant opens its platform to third-party sellers who bring the supply and benefit from the traffic; competition among them drives down prices and raises service quality, and the host "monetizes its infrastructure." This is the answer to the starting observation: traditional e-commerce isn't profitable because you're "always one click away from your competitor." Marketplace activity accounted for 5% of Amazon's business when Splitgames was founded; it now represents more than two-thirds.
2. What Fnac demonstrated. Under a strong brand with established e-commerce, the marketplace becomes the top source of profitability and the top acquisition channel, as well as a testing ground for categories that later move into stores. Competitors failed for lack of the dual know-how, technical and business — hence the idea of packaging it as software, paid on transactions just like the merchant itself: "our models are aligned."
3. The "SaaSpocalypse" broken down into two distinct risks. , not to be confused. (1) Per-seat pricing: "when you charge by number of users but you only have one user left, and that user is an AI, that can be a problem" — not an issue for transaction-based revenue, "whether it's an AI buying or a human customer, it makes no difference to us." (2) The rewriting of the product by a model. To be considered alongside [[greenwald-sierra-outcome-based-pricing-ai-agents-2024-12-10]].
4. The three arguments against rewriting. the asset isn't the code but an infrastructure of nearly 100,000 sellers, hence just as many information systems connected to those of its customers; Mirakl "has become a standard," and a standard is hard to rewrite; and finally, holding up at scale — $20 billion expected to flow through in 2026, billions of API calls, hundreds of millions of orders.
5. The claimed head start on AI. first AI team in 2017, products from the last five years built "100%" on AI, Mistral used alongside in-house models that perform better "on this type of specific task" — the argument being proprietary data, "a model without data is useless." An agentic e-commerce platform is under development.
6. The transformation method, reproducible as-is. Start from the deliverable — code, business reviews, support responses — then work back up the chain through reverse engineering: what's the best way to produce it with AI? Applied to support, the finding is that the common source of documentation, training, and support is the doc; the three teams, until then reporting to three different departments, are merged into a single one with an agent at the end of the chain. Stated result for OneHelp: at constant headcount, customer satisfaction ×2 and support users ×10. The claimed gain isn't headcount reduction but human availability freed up for the cases the agent doesn't handle — "the one thing AI won't be able to replace is the customer relationship."
7. The unexpected shape of attrition. no layoffs, but departures of employees saying "this isn't my job" and going to do it at a company "that does it the way we used to"; renewed motivation among those who stay; departing staff replaced.
8. The organizational thesis, stated twice. , backed by an argument about speed: "there's no adoption of a technology like AI without an organizational transformation," without which you optimize processes and miss the gains. Computing and the Internet took hold "over decades," AI "in years" — speed being, in his view, the real economic problem. To be compared with [[sfeir-adeo-decathlon-agent-bon-sens-2026-09-06]], whose Adeo group is cited here via Leroy Merlin.
9. The diagnosis on juniors and its deferred consequence. The current level of AI corresponds, in many professions, to "a junior's level"; juniors are therefore being replaced — and both speakers point to the same dead end: without juniors, "how are we going to train seniors and experts?" An open question acknowledged as such: is training juniors still the company's role, or a mix with higher education and sector-specific schools? The interviewer backs this with two references — Erik Brynjolfsson and the canaries in the coal mine (declining demand for young computer scientists in the United States, "correlation isn't causation yet"), David Autor on the democratization of expertise, the nurse taking on tasks once done by a doctor. Corrot agrees with the mechanism, but is cautious about the scale.
10. The usage study cited in closing. , directly useful for a training policy: among students, AI used to summarize, generate, or translate lowers performance; used to practice, generate exercises, and get corrected, it raises it.
11. The indictment of the French framework. , to be conveyed as an executive's stated position: people are trained for jobs known not to exist in two years; labor law blocks organizational change; France lacks a vision "perhaps even since De Gaulle," unlike China and the United States, which have one; Europe's vision amounts to regulation alone. On the model of society, he declines to take a side — "1984 or Brave New World, today we don't know" — and poses universal basic income as a question to be studied. To be read against [[dgtresor-ia-effets-emploi-2026-06-30]].