On January 6, 2026, Microsoft France unveiled the 15 startups selected for the 3rd edition of the Microsoft GenAI Studio at STATION F. Humind is one of them, in the "Commerce & Sales" category, for its e-commerce AI agents. Here is what the program puts on the table, how you get in, and why direct access to Microsoft's engineering matters so much when your conversational agents run every day on real stores, in front of real shoppers.
15 startups out of more than 180 applications
The Microsoft GenAI Studio is a four-month acceleration program run by Microsoft France and hosted at STATION F. For this third edition, the jury reviewed more than 180 applications before selecting 15, spread across health, education, finance, industry, customer relations, productivity and commerce. The list was made public in Microsoft France's press release of January 6, 2026, picked up the same week by J'aime les Startups.
Humind appears in the "Commerce & Sales" category. Microsoft describes the company as offering AI agents dedicated to e-commerce, optimizing customer interactions to facilitate sales and improve the shopping experience. The summary is accurate: an agent installed in the merchant's store, answering product questions, guiding discovery, recommending, and feeding the resulting conversational intelligence back to the merchant.
The edition's partners are STATION F, Mistral AI, NVIDIA, GitHub, Databricks, Cellenza and ENGIE. Eneric Lopez sums up the program's intent:
"With the launch of this third edition, we are accelerating the adoption of generative AI in the French economy, supporting startups that are inventing tomorrow's use cases. Our goal is to bring out technological champions and give them the concrete means to scale, responsibly and securely."
Eneric Lopez
Director of AI & Social Impact, Microsoft France
What the program actually puts on the table
The content announced by Microsoft is precise: up to $350,000 in Azure credits, access to more than 11,000 AI models through Azure AI Foundry, LinkedIn Premium, GitHub Enterprise and Microsoft 365 licenses, a dedicated workspace at STATION F, and mentoring workshops on responsible AI, data security, cost optimization and scaling.
The credits are the visible part
Cloud credits matter, but not for the reason you would imagine. They do not improve a product on their own. What they buy is the freedom to test seriously: running an evaluation across several model families instead of settling for the cheapest one, maintaining a pre-production environment that actually resembles production rather than a mock-up, reindexing an entire catalog three times in one week because the second attempt was bad. That freedom is usually rationed. For four months, it will not be.
The invisible part is the engineering time
Interviewed by SKEMA Business School a week after the announcement, Alexis called the selection a "true mark of trust", then pointed to what appears in no press release:
"Today, the Chief Architect of Microsoft France spends between four and five hours a week with us."
Alexis Hespelle
Co-founder, Humind, quoted by SKEMA Business School, January 14, 2026
Four to five hours a week with someone who has already seen hundreds of architectures in production are worth more than any credit line. It is the difference between discovering a scaling problem because a merchant calls you on a Saturday morning, and discovering it on a whiteboard in January.
And the whole ecosystem around it
The first weeks set the tone. The workshops organized by Microsoft have already brought us a lot, and a dinner with NVIDIA and the founders of the other startups in the cohort was a reminder of what makes this program special: you do not just benefit from Microsoft, but from the entire ecosystem gathered around it, partners, engineers and founders solving neighboring problems on other markets.
Why infrastructure matters as much as the model for an e-commerce AI agent
The idea that an AI product comes down to the choice of model does not survive contact with production. Once an agent is live on a real store, most of what decides whether shoppers trust it has little to do with the model that answered.
A conversation is not a demo
A demo agent answers a question. A production agent handles thousands a day, across time zones, in eleven languages, on catalogs that change overnight. It has to stay consistent while a catalog sync is running, degrade cleanly when a third-party API slows down, and never offer a shopper a product that went out of stock four minutes earlier. None of that is solved in a prompt. It is a matter of regions, caching, retries, queues, observability and plumbing that has to be testable.
Trust is a property of the infrastructure
Merchants who put an agent in front of their own customers ask infrastructure questions well before model questions: where is the data, who can read it, what happens in an outage, how do you prove what the agent said. Building on Azure lets us answer in terms a CIO recognizes: managed identities rather than keys that never expire, regional deployments, content filtering, and a telemetry trace behind every request. Those answers are written in our trust center, not in a sales deck.
Cost per conversation is a product decision
Every conversation has a cost, and that cost silently decides what the product can afford. Whether an agent can re-read a long product page, run a second search pass or verify its own answer is as much an economic question as a technical one. Model routing, caching, search quality and prompt design are all cost levers. The workshops on cost optimization and scaling are, on paper, the ones we are looking forward to most.
A journey, not a lucky break
The January 2026 selection did not come out of nowhere. Humind joined STATION F's Founders Program at the end of 2024, right after a stint at Berkeley SkyDeck, the University of California's accelerator where the two founders, Alexis Hespelle and Dan Castiel, both SKEMA graduates, laid the foundations of the product.
We already knew the GenAI Studio: we had applied to a previous edition, in March 2025. During the technical assessment week, the Microsoft team advised us to let the product mature and submit our application again the following year. That is exactly what we did. In the meantime, the architecture evolved, the agent became measurably better at its job, and the answers we could give under technical scrutiny changed with it. We reapplied, and this time we were selected.
See you in May
The cohort will close with a Demo Day in Paris in the spring, where each team will present the road covered. Until then, four months of concentrated work await us, with people who have already solved the problems we are about to meet, and the room to fix things properly rather than quickly. For the merchants running our agentic commerce agents, that is what matters.
To see what an e-commerce AI agent looks like on a real catalog rather than in a demo video, browse our e-commerce use cases or request a demo.
