AI in France

ThinkDeep AI built DeepBrain, a sovereign multi-agent platform using NVIDIA AI Enterprise, NIM microservices, NeMo Retriever, NeMo Guardrails, and DGX H100/H200 hardware, cutting French government document processing from 2 days to 2 minutes and saving €2M at the Ministry of Economy and Finance at 10,000-employee scale.

Rakuten France runs 80+ Custom Agents in Notion, most built by employees outside its four-person AI team. AI Lead Sidney Golstein wrote shared operating procedures and a meta-agent, the Rakuten Agent Builder, that turns a plain-language description into a working agent in 10 to 20 minutes, plus an Agent Registry, run logs and an improver agent that proposes fixes. Some product teams cut backlog health issues by 50%.

Sanofi's Quality Regulatory Science team is scaling Plai, an in-house AI quality-intelligence tool, from a deviation-trending pilot to a global deployment across multiple manufacturing sites. Built on the RAISE (Responsible AI at Sanofi Enterprise) governance framework, Plai accelerates investigations, generates root-cause hypotheses, and integrates QMI and QRE risk metrics to surface predictive insights before deviations occur.

ThinkDeep AI built DeepBrain, an on-premises multi-agent platform on NVIDIA DGX and NIM, for the French Ministry of Economy and Finance and Ministry of Defense. Rolled out to 10,000 finance ministry staff, it saved €2M and cut legal document retrieval from up to two days to two minutes.

Paris-based multiplayer AI platform Dust deployed Claude via Anthropic’s API to power its no-code AI agent infrastructure, saving roughly $10,000 per day on model spend (18-19% reduction) after optimizing prompt caching. The deployment raised autonomous tool-calling depth from 4-5 to 8+ steps per agent run with no engineering changes and now serves more than 300,000 agents deployed across customers including Datadog, Vanta, and 1Password.

Leboncoin deployed a multi-agent LLM customer-service system handling 60,000 emails/week with 90% AI-automation, freeing 800 working hours/month and reducing median response time from 8 to 2 hours for France's leading classifieds platform.

Stripe Systems (Paris/India fintech engineering firm) deployed AI-assisted code review via GitHub Actions across 3 production projects (NestJS, Flutter, React) and cut review cycle time 60% (4.2h → 1.7h), senior engineer review time 60% (45 → 18 min), and increased defects caught per 100 PRs 35% (8.2 → 11.1). False positive rate dropped 22% → 8% over 3 months via confidence threshold tuning + 👍/👎 feedback loop. $80-120/month token cost for ~45 PRs/week.

Verkor's Design Conductor AI agent autonomously generated a layout-ready GDSII RISC-V CPU (VerCore) in 12 hours from a 219-word spec, clocked 1.48 GHz with CoreMark 3,261. Verified in simulation only — not yet fabricated on silicon; 5-10 human experts still needed for production-grade quality.