Choco uses OpenAI to automate food distribution and double sales productivity

Choco has shown how AI agents are moving from auxiliary tools to critical infrastructure in a real business. The food distribution platform built the OpenAI API into its order processing and, by its own figures, now runs more than 8.8 million orders a year through the system, cuts manual entry by up to 50% and lets sales teams work twice as productively without growing headcount.

Choco operates in food & beverage distribution and brings restaurants, suppliers and distributors together in one system. The company says it serves more than 21,000 distributors and 100,000 buyers in the US, the UK, Europe and the Gulf states.

As the business grew, the main problem turned out not to be interfaces, or the CRM as such, but a chaotic inbound stream of orders: they arrived by email, SMS, voice messages, photos of documents and even handwritten notes. Staff then converted all of this by hand into structured orders for ERP systems, which slowed scaling and created a constant risk of errors.

According to the company, the hardest part was not even recognizing text or speech. The critical layer is hidden context: each customer's own SKU mappings, usual units of measure, delivery patterns and other details that typically live in the heads of order desk employees. That is why Choco bet not simply on automating templated steps, but on an inference system that can resolve ambiguity the moment an order arrives. In effect, it moves operational knowledge out of human memory and into a software layer that is available 24/7 and does not depend on a particular shift or employee.

On that foundation the company launched OrderAgent, an agent for handling multimodal inputs that accepts emails, SMS, images and documents and turns them into ERP-ready orders. VoiceAgent followed, built on the Realtime API: it lets customers place orders by phone with latency under a second, including outside business hours.

Choco says it chose OpenAI for model quality, multimodality, structured outputs and production reliability. Technically, the rollout involved speech-to-text, embeddings and function calling, and on top of that the team built its own quality evaluation loop: reference datasets, continuous monitoring and A/B tests. This detail matters: in systems like these, it is not enough to "plug in an LLM"; you have to keep measuring where the model goes wrong and how quality changes on real orders.

The results read less like an experiment and more like industrial operation. Choco claims that more than 200 billion tokens pass through OpenAI in production, order intake runs around the clock, and the error rate stays low with configurable automation thresholds. Cutting manual entry by up to 50% frees employees from mechanical processing and moves them to higher-value tasks, while a twofold gain in sales productivity without hiring shows the effect reaches beyond operating costs to commercial efficiency.

The UX side matters too: customers did not have to change their habits. They still send orders in whatever way suits them, and it is the system that adapts.

Choco also highlights several management lessons. First, build quality evaluation from day one, even if the team has only 10–20 labeled examples. Second, AI systems need their own observability: ordinary logs are not enough if you cannot see the model's inputs, outputs and the reasons for failed runs. Third, a business needs to accept the probabilistic nature of LLMs up front. Such systems do not behave like deterministic code, so user expectations, SLAs and escalation paths have to be designed differently.

This is a fairly mature signal for the market: the winners will not be those who added a chatbot fastest, but those who learned to build managed agent workflows around complex business data.

The main takeaway from the Choco case is simple: AI agents are becoming a new operational layer in vertical B2B, where value comes not from a polished interface but from the ability to parse a messy real-world data stream and turn it into action. If this model scales to other supply chain segments, the market will see not spot automation but the gradual replacement of entire manual functions by systems that can listen, read, clarify context and do the work themselves.