Choco burned through 200 billion AI tokens to solve a problem most tech companies never think about: who's going to answer the phone at 2 AM when restaurants need food?
The Berlin-based food distributor just hit this milestone processing real orders across 110,000 businesses. Not chatbot demos. Not synthetic benchmarks. Actual purchase orders for tomatoes and steaks flowing through OpenAI's APIs.
Daniel Khachab and Julian Hammer founded Choco in 2018 targeting the most unsexy corner of tech: food distribution. Smart move. While everyone chased consumer apps, they built infrastructure for an industry that moves billions in inventory daily.
<> The flagship product is the Choco VoiceAgent, launched in December 2025, which operates on OpenAI's Realtime API and can receive calls, take orders, answer product questions, and recommend items in any language, 24/7./>
That December 2025 launch date caught my attention. Either the research is from the future, or Choco's timeline got scrambled somewhere. Regardless, the technical approach is what matters.
Night Shift Economics Nobody Discusses
Here's what the press releases won't tell you: food distributors can't staff night shifts. High turnover. Constant retraining. Answering machines from 1995.
Restaurants place orders after service ends—usually 10 PM to 6 AM. Someone needs to process those orders, check inventory, suggest alternatives. Humans hate this work. Choco's AI agent doesn't.
The numbers actually make sense:
- 95% order accuracy (better than tired humans at midnight)
- 50% reduction in manual processing time
- Zero sick days or vacation requests
What Nobody Is Talking About
Choco's real innovation isn't the voice interface—it's the ERP integration with operational guardrails. Anyone can build a chatbot. Few can make it check actual inventory, propose alternatives for out-of-stock items, and push promotions on food nearing expiration.
They prioritized prompt engineering over fine-tuning. Faster iteration. Model flexibility. This is how you actually ship AI products instead of endless training experiments.
The technical stack reveals their pragmatism:
- Few-shot learning with dynamically retrieved examples
- Semantic embedding-based retrieval for messy input formats
- Context injection handling PDFs, voicemails, text messages
No fancy research papers. Just boring infrastructure that works.
The $328 million unicorn valuation suddenly makes sense when you realize they're automating manual workflows across an entire industry. Every token processed represents work a human used to do manually.
Infrastructure Play, Not Consumer Gimmick
Choco demonstrates enterprise AI done right. They didn't build a general-purpose assistant hoping someone would find a use case. They identified a specific operational pain point and built AI to solve it.
Food waste reduction through real-time stock confirmation? Revenue increase via expiration promotions? These aren't AI features—they're business logic powered by AI.
The voice agent handles multiple languages automatically. Try finding bilingual night shift workers in most cities. Good luck with that hiring process.
My take: This is what enterprise AI adoption actually looks like. Not flashy demos or chatbot integrations. Replacing expensive, hard-to-staff manual processes with systems that work reliably at scale.
Choco's 200 billion tokens represent something more valuable than training data—they represent production workload in a massive industry. While everyone builds RAG chatbots, they built mission-critical infrastructure.
Smart positioning. Boring execution. Profitable outcomes.
