Choco automates food distribution with AI agents
Key Points
- Processes 8.8M+ orders/year
- OrderAgent handles multimodal inputs
- Up to 50% reduction in manual entry
Summary
Choco integrated OpenAI APIs to automate order capture and execution across global food distribution. They built multimodal OrderAgent (emails, SMS, images, documents) and a low-latency VoiceAgent (Realtime API) to convert unstructured inputs into ERP-ready orders, applying in-context customer mappings and catalog logic. The platform processes 8.8M+ orders annually, has processed 200B+ AI tokens in production, reduces manual order entry by up to 50%, and doubles sales productivity without added headcount while maintaining error rates in the ~1–5% range.
Key Points
- Core components: OrderAgent (multimodal parsing + structured outputs) and VoiceAgent (Realtime speech-to-text + sub-second latency).
- Key capabilities used: speech-to-text, embeddings, function calling, SDKs, and unified multimodal model handling.
- Customer-specific inference: dynamic in-context learning to resolve SKU mappings, unit preferences, and delivery patterns at capture time.
- Evaluation & reliability: ground-truth datasets (even 10–20 examples), continuous monitoring, A/B testing, and AI-native observability capturing inputs, outputs, and reasoning traces.
- Operational impact: supports 24/7 ordering, eliminates millions of manual workflows, achieves up to 50% manual-entry reduction, and enables scaling without headcount growth.
Practical notes for engineers
- Prototype with a small labeled dataset to measure progress and iterate quickly.
- Log model inputs/outputs and reasoning traces to enable debugging and continual improvement.
- Use embeddings + function calling for stable, structured ERP outputs and maintain configurable automation thresholds to control error rates.
- Validate latency and reliability for realtime voice flows under expected traffic to preserve sub-second experiences.