Scaling GenAI at Mirakl: From Catalog Transformer to AI-first Product Organization

How Mirakl reduced onboarding from 11 days to hours.

Mirakl has successfully pivoted its product strategy by integrating generative AI to automate complex marketplace operations, most notably through its Catalog Transformer. By transitioning from a project-based data science model to a permanent “Squad” organization, the company has drastically reduced customer onboarding times while managing the high costs of LLM deployment. This strategic shift positions AI as a core architectural pillar rather than a peripheral feature, enabling significant scalability for global retailers.

Points clés

Mirakl is a French tech scale-up founded 12 years ago by Philippe Corrot and Adrien Nussenbaum, currently employing 750 people globally.
The company powers over 450 customers and manages $8 billion in gross merchandise volume (GMV).
Early AI implementations like “Catalog Auto-mapping” and “Customer Care Intelligence” reduced customer incidents by 90% before the generative AI boom.
The introduction of the Catalog Transformer shifted manual vendor onboarding from a 15-day process to just two hours using GenAI.
Initial tests with GPT-4 for catalog processing projected a non-sustainable ROI cost of €500,000 per month.
To optimize costs and performance, Mirakl implemented a layered approach using fine-tuned smaller models alongside larger LLMs.
The data science team, known as “MiraDoge,” scaled from 7 members in 2020 to 40 members today to support an AI-first product roadmap.
Mirakl transitioned its organizational structure from isolated AI projects to a “Quattro” model, integrating product, design, engineering, and data science.
The company now operates four dedicated AI squads focused on specific product areas like Marketplace Platform and Retail Media (Ads).
Dedicated AI research engineers are now utilized to investigate complex theoretical problems months before product integration.

À retenir

If you want to stay relevant, just listen to your co-founder when he starts playing with ChatGPT and realizes your years of hard work can be mimicked by a prompt—it’s a great motivator for “disruption.” Mirakl’s secret sauce seems to be spending half a million euros a month on API calls just to prove a point, before realizing that maybe, just maybe, smaller models are better for the wallet. But hey, at least they traded 11 days of human boredom for two hours of machine magic, proving that if you throw enough “Doge” data scientists at a problem, you might actually ship something useful once in a decade.

Sources