Choco: order capture needs customer context, not just transcription

External-source research synthesis, not an OfferSignal or Sander Gansen client result. This edition consolidates the retained October 2026 research, including the 23 and 27 October horizon; source-specific dates and evidence limitations remain distinct.

A food distributor may receive the same repeat customer's order by email, SMS, voicemail, photograph, PDF or handwritten note. Accurate transcription does not establish whether shorthand means an individual item, a pack or a crate. The distributor still needs customer identity, the right stock-keeping unit (SKU), customer-specific price, substitution permission and a credible delivery window. Choco's case is therefore about governed order capture, not simply converting speech or text into an order.

From restaurant messaging to distributor automation

Choco began as a mobile ordering and messaging layer connecting restaurants and suppliers. In 2019 it described an app replacing recurring voicemail and fax orders, with a reported presence in 15 cities across Europe and the United States. During the 2020 pandemic it tested direct-to-consumer restaurant-quality produce websites in 17 markets; the retained reporting does not establish a lasting product line. 1 2

The retained bibliography also includes a $100 million Series B announcement and Choco's unicorn-status announcement. These are historical financing context, not evidence of AI order accuracy or customer productivity. 3 4

The current distributor-side proposition is OrderAgent: accepting email, SMS, images and documents, extracting order details, mapping customer-specific catalogue terms and producing structured, ERP-ready orders. VoiceAgent addresses telephone ordering through OpenAI's Realtime API. Customer-specific SKU mappings, units and delivery history matter alongside transcription. The OpenAI case describes ground-truth datasets, monitoring, A/B testing, configurable confidence thresholds and human review for edge cases. Optional Autopilot is presented as bounded execution, not unrestricted autonomy. These descriptions do not establish that every telephone interaction can safely release an order. 6

Keep scale, productivity and accuracy separate

The 2026 OpenAI customer case reports more than 8.8 million orders processed annually, up to 50% less manual order entry and 2x sales-team productivity without added headcount. It also reports 21,000 distributors and 100,000 buyers across the US, UK, Europe and the Gulf Cooperation Council. These are provider/OpenAI case-study claims. Annual order volume describes scale, not profitable new demand. The publication supplies no cohort, baseline period, geography split, causal design or independent audit for the productivity claims, so they are not universal productivity or return-on-investment results. 6

A named Colony Foods account says that after one year of OrderAgent, the Massachusetts distributor saved at least two hours per day and reduced order-entry errors. This is customer-reported evidence without an order denominator, before-and-after sample, control or independent audit. Separately, a 2020 trade report described one French supplier's Broadcast pilot with a 100% open rate and 53% purchase conversion. Its sample size, period and conversion definition are undisclosed; this earlier messaging result is not an AI order-entry accuracy measure. 9 5

Current Choco and Dot Foods materials also cite 1m+ monthly orders, 99% AI order accuracy and 20% larger average order value. Keep these claims separate from the annual volume and productivity figures: the retained material does not establish a shared population or measurement period. The 99% figure is a partner marketing claim without a published test population, not demonstrated exact-line, quantity, unit, price or margin accuracy. A configurable confidence threshold is likewise not a universal accuracy guarantee. 6 10

The older note records a publicly accessible help-centre correction video. No identifiable public URL was retained, so this remains an unverified observation, not an independently checkable demonstration. Even taken at face value, a correction interface does not establish an order accepted through a live customer's ERP, voice-channel reliability or permission to release an order.

The broader suite has a boundary

Choco also positions captured order data for branded ecommerce, sales prospecting, quoting support and customer-health signals. Its 2024 Sales Rep App added alerts, catalogue access, screenshot-to-order processing and menu scanning. A Dot Foods partnership extends the model to more than 4,000 Dot distributor customers, with ERP connection and an Endless Aisle of 42,000 Dot SKUs. Partnership reach and catalogue breadth are not measured adoption or successful order processing by every eligible customer. 7 8

This is a useful distributor data-and-workflow pattern. It is not evidence of an engineered industrial request-for-quotation or configure-price-quote system handling drawings, revisions, tolerances, materials, tooling or production capacity. Custom drawing-based manufacturing still requires engineering and feasibility decisions beyond the food-ordering pattern.

A proposed pilot: useful drafts before write access

For an illustrative packaging or fastener distributor, begin in shadow mode, not automatic order release. Map customer account, authorised requester, customer alias, seller SKU, pack-versus-each quantity, price-list version, tax basis, stock or available-to-promise, delivery history and window, substitution permission, credit limit, credit hold and credit authority. Compare extracted lines with the original message and an approved ground-truth set. A representative set of 10–20 orders can start a benchmark, but a production sample is needed before broader write access. This is a proposed evaluation approach, not a reported Choco deployment result.

Flag unexpected units, unusual quantities, expired prices and unapproved replacements. The order-desk or commercial owner—not the model or inbound message—remains authoritative for new SKUs, nonstandard conditions, margin deviations, delivery changes, credit holds and substitutions. Assign a named reviewer. A draft can be useful without permission to release an order; telephone capture does not remove that distinction.

Preserve the source message, extracted fields, confidence, rule version, timestamp, reviewer correction, ERP response and acceptance acknowledgement. An apparently successful extraction is not proof of downstream acceptance. Include idempotency checks, role-based permissions, confidence thresholds and rollback before any Autopilot action.

Measure exact-line, quantity, unit and price accuracy; review minutes; duplicate and missed orders; straight-through rate; cycle time; margin leakage; complaints and customer escalations. Split results by customer, input channel, item complexity and exception type. Accuracy on neat email orders may not transfer to unfamiliar handwritten aliases or voice orders. The operating owner must be able to identify failures, correct mappings and establish what the downstream system actually accepted.

Commercial fit and unresolved questions

Choco's strongest evidenced fit in the retained material is messy multi-channel order capture into ERP, followed by distributor ecommerce, sales and customer workflows. Dot customers are offered preferred support, discounted pricing and free ERP integration; current public numeric pricing and contract terms remain undisclosed. Budget for catalogue cleanup, ERP mapping, onboarding, reviewer labour, monitoring and rollback rather than assuming the entry-reduction figure is a complete business case. 10

The evidence supports a credible operating B2B food-distribution vendor, not controlled proof of universal 2x productivity, 99% accuracy or ROI. Confirm the supported country and language matrix, exact ERP connectors, data-retention terms, exports and commercial packaging in a distributor demo. The lesson is to govern commercial facts and verify acceptance before granting an agent permission to act.

Sources and next reading

The distributor platform provides a public overview, with demo/request access for commercial use. The restaurant app remains separately listed on Apple App Store and Google Play; availability and regional terms may vary. Dot's partner page describes product information and customer commercial access and discounts. These access routes are not verification of a live customer order. 11 12 13 10

Explore OfferSignal · Focused Business & AI Workshop · All research

References