Choco Revolutionizes Food Logistics with OpenAI's AI

Choco and AI: A New Era for Food Distribution

Choco, an innovative platform powered by artificial intelligence, is redefining the distribution of food and beverage products. By connecting over 21,000 distributors and 100,000 buyers across the United States, the United Kingdom, Europe, and the GCC, Choco simplifies the management of orders, sales, and customer relationships throughout the supply chain.

In the face of increasing order volumes, Choco had to overcome a major challenge: the diversity of order formats, ranging from emails to handwritten notes. This manual and error-prone process limited the company's efficiency. The real issue lay in the implicit context, which included customer-specific SKU mappings, unit preferences, and delivery patterns. This knowledge was held by order service representatives and needed to be encoded into inference layers to resolve ambiguity during order capture.

“Processing these inputs was the first hurdle, but not the most difficult. The real problem was the implicit context: customer-specific SKU mappings, unit preferences, delivery patterns. This knowledge was in the heads of the order service representatives, and we needed to encode it into inference layers that resolve ambiguity at the moment of order capture.” — Narbeh Mirzaei, VP Engineering

With the emergence of production-ready large language models (LLMs), Choco saw an opportunity to surpass traditional workflow software. The integration of OpenAI APIs became central to this transformation, enabling Choco to build AI systems capable of executing work directly.

Integration of OpenAI APIs

Choco integrated OpenAI APIs to create AI-native products, such as OrderAgent, which converts multimodal orders into structured data ready for ERPs. VoiceAgent, also based on OpenAI, allows for phone orders with latency under one second.

OpenAI was chosen for its multimodal capabilities and reliability. This integration allowed Choco to unify previously disconnected workflows, thereby optimizing operational efficiency. Choco built a dynamic contextual learning infrastructure to resolve ambiguity based on each client's order history and catalog, distinguishing automation from intelligence.

The implementation was rapid and scalable. By using OpenAI's SDKs and APIs, Choco quickly integrated capabilities such as speech-to-text conversion, embeddings, and function calling into its infrastructure. The team also built a rigorous evaluation framework with ground truth datasets, continuous monitoring, and A/B testing to ensure accuracy and performance in production.

Adoption was facilitated by seamless integration throughout the order workflow. Customers did not need to change their ordering methods—whether by phone, text, or email, the system adapted to them.

Impressive Results

Thanks to this technology, Choco processes over 8.8 million orders per year, reducing manual entry by 50% and doubling productivity without increasing headcount. Error rates are maintained between 1% and 5%, with order availability 24/7.

Choco also developed VoiceAgent, powered by OpenAI's Realtime API, allowing customers to place orders naturally over the phone with latency under one second—even outside of business hours.

Leadership Strategies and Future

Choco emphasizes continuous evaluation and AI-native observability to enhance its systems. The company plans to expand its AI capabilities to strengthen the autonomy and contextual awareness of systems across sales and the supply chain.

By leveraging OpenAI APIs, Choco continues its transformation from a workflow platform to an AI-powered execution infrastructure, paving the way for a new class of users capable of managing intelligent systems.

Leadership Lessons

Choco has learned the importance of starting with an evaluation from day one. Even a small set of ground truth data (10–20 examples) allows teams to measure progress, validate improvements, and iterate with confidence.

Investing in AI-native observability is crucial. Debugging AI systems requires more than traditional logging—capturing inputs, outputs, and model reasoning traces is essential for understanding and improving performance.

Setting the right expectations from the outset is also key. Unlike deterministic software, LLMs are probabilistic. Educating teams and users about this difference is essential for building trust and avoiding friction during adoption.