Copilots are 20% of AI’s value. Michelin is chasing the other 80: Inside TFN’s day at HumanX Amsterdam

At HumanX Amsterdam, Tech Funding News spent the day circling one stubborn theme: companies are still better at demoing AI than deploying it. That showed up in a panel Akansha Dimri moderated with Michelin’s Ambica Rajagopal and causaLens CEO Darko Matovski, and again later in the pitch competition, where six early-stage startups tried to prove they were building something more durable than a slick wrapper.

Rajagopal’s argument was blunt. Michelin already has more than 30,000 people using its internal generative AI platform, but she said that is only about a fifth of AI’s value to the company. The bigger return comes from custom models tied to Michelin’s own processes — manufacturing forecasts, tyre design, and the company’s much messier planning work. Over three years, she said, AI has delivered more than $200M in value and grown at around 30% a year. She also drew a hard line between consumer AI and enterprise software: a ChatGPT-style error rate inside Michelin would have been a problem, not a feature.

Matovski made the same point from the other side of the table. Copilots, he said, make individuals more productive. Digital workers do the work. That matters because the gains from a copilot are limited, while the token costs can swing wildly. His example was demand forecasting for each SKU, which can involve around 20 teams and a pile of software. You can speed people up a bit, but you don’t change the process until something cuts across it.

He also said the main blocker is not technical. It is leadership. When causaLens gets a mandate to redesign a process, it succeeds. When it is boxed in, it doesn’t. Rajagopal’s own answer at Michelin was to build rather than buy AI agents two years ago, even if that meant spending more upfront. Her logic was simple: paying forever for someone else’s commodity was not the play.

The startup pitches later in the day reflected the same shift. HumanX limited entry to companies that had raised under €10M, been operating for less than five years, and had something real to show. The semi-finalists included LangWatch, 8wave, Avendar, Hola AI, Weeve and Methodino — all aiming at infrastructure, governance, sovereign systems, private assistants or confidence scoring, not just chat. That is the market now: less “look what the model can say,” more “can this survive contact with a company.”