Choco automates food distribution with AI agents

Source: OpenAI Blog

The Bottleneck

As order volumes grew, Choco hit a major bottleneck: orders still arrived through emails, texts, voicemails, images, and even handwritten notes. Teams manually translated those inputs into structured ERP orders—a slow, error‑prone process that limited scale and created constant operational friction.

“Processing those inputs was the first barrier, but not the hardest one. The real problem was implicit context: customer‑specific SKU mappings, unit preferences, delivery patterns. That knowledge lived in the heads of order desk reps, and we needed to encode it into inference layers that resolve ambiguity at the point of order capture.”
— Narbeh Mirzaei, VP Engineering

AI‑Native Products

OrderAgent

Choco embedded OpenAI APIs at the core of its platform to power a new generation of AI‑native products. The company introduced OrderAgent, which processes multimodal inputs—including emails, SMS, images, and documents—and converts them into structured, ERP‑ready orders.

“The transcription and extraction capabilities gave us a strong foundation. The real engineering challenge was building dynamic in‑context learning infrastructure, so the system resolves ambiguity against each customer’s ordering history and catalog. That’s what separates automation from intelligence.”
— Narbeh Mirzaei, VP Engineering

VoiceAgent

Choco also built VoiceAgent, powered by OpenAI’s Realtime API, enabling customers to place orders naturally over the phone with sub‑second latency—even outside business hours.

Implementation Highlights

• OpenAI was selected for its model performance, multimodal capabilities, structured outputs, and production reliability at scale.
• Using OpenAI’s SDKs and APIs, Choco rapidly integrated speech‑to‑text, embeddings, and function calling.
• A rigorous evaluation framework with ground‑truth datasets, continuous monitoring, and A/B testing ensures accuracy and performance in production.

Business Impact

• Processes over 8.8 million orders annually, eliminating millions of manual workflows.
• Achieves up to 50 % reduction in manual order entry, freeing teams for higher‑value work.
• Enables 2× productivity gains, allowing teams to scale without additional headcount.
• Maintains error rates below 1–5 % with configurable automation thresholds.
• Supports 24/7 order intake, eliminating delays from nights and weekends.

Key Learnings

1. Start with evaluation from day one – Even a small ground‑truth dataset (10–20 examples) enables teams to measure progress, validate improvements, and iterate with confidence.
2. Invest in AI‑native observability – Debugging AI systems requires more than traditional logs; capturing model inputs, outputs, and reasoning traces is essential to understand and improve performance.
3. Set the right expectations early – Unlike deterministic software, LLMs are probabilistic. Educating teams and users on this difference is key to building trust and avoiding friction during adoption.

Future Roadmap

Choco is continuing to expand its AI capabilities across the food distribution ecosystem, deepening the role of agents in executing complex operational workflows. The company is enabling a new class of users—non‑engineers who act as “agent orchestrators,” designing and managing intelligent systems that drive business outcomes.

Looking ahead, Choco plans to further scale its use of OpenAI APIs to power more autonomous, context‑aware systems across sales, commerce, and supply‑chain operations—continuing its shift from workflow software to AI‑powered execution infrastructure.