Choco built an inference layer to encode order-desk knowledge, not just transcribe orders

The bottleneck was messy inputs; the real problem was tacit knowledge

Choco connects restaurants, suppliers, and distributors — over 21,000 distributors and 100,000 buyers across the US, UK, Europe, and the GCC. As volumes grew, orders kept arriving through emails, texts, voicemails, images, and handwritten notes that staff manually retyped into structured ERP orders.

The company is explicit that ingesting those formats was the easy part. What blocked scale was knowledge that lived in people's heads: which SKU a given customer means, their preferred units, their delivery patterns.

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.Montana Labs

This is a useful distinction for anyone building extraction systems. Transcription and OCR are commodity capabilities now. The differentiator Choco names is resolving ambiguity against each customer's ordering history and catalog — what VP Engineering Narbeh Mirzaei calls the line separating automation from intelligence.

Two agents, one multimodal stack

Choco built OrderAgent to convert multimodal inputs — emails, SMS, images, documents — into structured, ERP-ready orders. VoiceAgent, built on OpenAI's Realtime API, lets customers place orders over the phone with sub-second latency, including outside business hours.

The stated reason for consolidating on one vendor was practical: handling text, vision, and audio in a single ecosystem let Choco unify previously disconnected workflows. The team integrated speech-to-text, embeddings, and function calling via OpenAI's SDKs, rather than stitching together separate specialist providers per modality.

The reported results are concrete: 8.8M+ orders processed annually, 200B+ tokens in production, up to a 50% reduction in manual order entry, and 2x sales-team productivity without added headcount.

Confidence thresholds and evaluation did the trust work

The design choice that stands out is Autopilot — an optional mode that automates processing only when confidence thresholds are met, routing edge cases to human review. Choco reports error rates below 1–5% with configurable automation thresholds, and a system that learns from corrections over time.

That gating matters because LLM order entry is probabilistic, not deterministic. Choco pairs it with a rigorous evaluation framework: ground-truth datasets, continuous monitoring, and A/B testing. Its own leadership lesson is to start with even 10–20 ground-truth examples on day one, and to invest in observability that captures model inputs, outputs, and reasoning traces — not just conventional logs.

Once customers saw it working with their own orders, trust followed quickly. That's when adoption really accelerated.Montana Labs

Adoption also benefited from not asking customers to change behavior. Buyers kept ordering by phone, text, or email; the system adapted to them rather than the reverse.

What Choco's rollout signals: agent orchestration as a non-engineer job

The forward-looking claim in this announcement is organizational, not technical. Choco says it is enabling a new class of users — non-engineers who act as 'agent orchestrators,' designing and managing intelligent systems that drive business outcomes.

Read alongside CEO Daniel Khachab's framing of moving 'from software that supports work to systems that actually do the work,' the specific implication is that Choco is treating order capture as a task to be executed autonomously under human-set confidence policies — with the order-desk rep's role shifting from data entry to defining thresholds, reviewing exceptions, and correcting the system. Whether that role holds up as agents take on more of the supply-chain workflow is the real test ahead.