Industries / Retail, hospitality and personal services / Wholesale trade / Foodservice distribution
Ordering
Ordering is how a kitchen tells its distributor what it needs, through a sales representative, a phone call, a fax, an app or a message between computers, and how that request becomes a list of exact products and pack sizes.
4
Milestones achieved, from 2000 to 2014.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2000
Six cases in ten barcoded
In the late 1990s ten trade associations, including the International Foodservice Distributors Association (IFDA) and the International Foodservice Manufacturers Association (IFMA), started Efficient Foodservice Response (EFR), a program to agree on product codes and electronic ordering between makers, distributors and operators. An EFR study presented in 2003 found bar codes on 54% of the food cases inspected in 1999 and 61% in 2000, but on only 11% of produce cases. Industry experts valued the savings from adopting EFR at $14 billion to $20 billion, Restaurant Business reported.
LessonCodes spread slowly while each link waited for the next one: makers said customers had not asked for bar codes, and customers waited for makers to print them.
Credit
the EFR sponsoring associations; Mary Gearing of Rich Products presented the study
2003
Japan's food trade orders online
Infomart, founded in Tokyo in February 1998, launched an online ordering system in February 2003, now called BtoB Platform Ordering. Restaurant chains, independent restaurants and hotels use it to order from food wholesalers, and wholesalers use it to keep orders that arrive by phone or fax as data. Infomart still sells an AI-OCR service (software that reads scanned text) that turns faxed purchase orders into data. Infomart counts ¥71 trillion of transactions on its platform in fiscal 2025, across all its services and not only food.
LessonOnline ordering did not end ordering by fax; the platform added a tool to read the faxes.
Credit
Infomart
2009
One number for every case
In October 2009, 55 foodservice manufacturers, distributors and restaurant operators launched the Foodservice GS1 US Standards Initiative to adopt GS1 standards: a Global Trade Item Number (GTIN) for each product, a Global Location Number (GLN) for each site, and product descriptions shared through the Global Data Synchronization Network (GDSN), an electronic network for product data. The initiative, endorsed by IFDA, IFMA and the National Restaurant Association, aimed for 75% adoption, measured by revenue, by 2015. Two years in, GS1 US said the industry had passed the halfway mark and foodservice GDSN subscribers had grown from 191 in 2009 to 1,430. Mike Roach of Ben E. Keith Foods said that before the initiative 6,000 distributors each moved makers' data into their own systems and the data "wasn't always accurate".
Credit
GS1 US; founding members including Sysco, U.S. Foodservice and Martin Bros. Distributing
2014
An app for small kitchens
Meicai ("beautiful vegetable"), founded in Beijing in 2014 by Liu Chuanjun, set out to supply about 10 million small and medium restaurants in China, whose owners order vegetables from a smartphone app. It raised $450 million in January 2018 at a valuation of about $2.8 billion, according to people familiar with the deal, as Bloomberg reported. After a push into group buying for households, it returned to restaurants in early 2021, shelved a US share listing in July 2021, and in September 2021 closed its research center in Chengdu and planned to cut at least half the staff of several Beijing teams, TechNode reported. Kuailv, the restaurant supply arm of its rival Meituan, also cut back its services in August 2021.
LessonMeicai's move into selling to households ended within a year and sent it back to restaurants.
Credit
Liu Chuanjun and Meicai
Watch for
AI enters phoned-in orders
Choco, a software company serving food distributors in the US, UK, Europe and the Gulf states, says orders still reach distributors as voicemails, texts, WhatsApp messages, emails and photos of handwritten notes. In March 2025 it said its Autopilot agent decides for itself whether an incoming order can go straight into the distributor's order system or needs a person, and that it already processed 50% of orders for early users such as The Menu Partners. In December 2025 it launched a Voice Agent, built on OpenAI's Realtime API, that takes orders by phone at night. An OpenAI customer story says Choco processes more than 8.8 million orders a year for over 21,000 distributors. Choco has published no error rates.
Achieved whena distributor or an outside tester publishes, for a stated period, the share of orders entered with no person involved and the share that later had to be corrected.
Claimed by
Choco, March and December 2025
Menus turned into order guides
An order guide is the list of products, pack sizes and prices that a sales representative prepares for a restaurant to order from. US Foods says its Automated Order Guide, built with Anthropic's Claude on Amazon Bedrock and with Amazon Textract to read paper menus, spreadsheets and handwritten notes, is used by 3,300 sellers, cut the time per proposal from 3 to 4 hours to 20 minutes and saved over 32,000 hours in six months. GrubMarket announced a Sales AI Agent on 8 June 2026 that reads photos or PDFs of menus and matches what it finds to a distributor's catalog. Neither has published how often the suggested products matched what an experienced representative would choose or what the restaurant then bought.
Achieved whenan outside party compares the drafted guides with guides built by experienced representatives, or with what restaurants ordered, and publishes the match rate.
Claimed by
US Foods, in an Amazon Web Services case study; GrubMarket, June 2026
Open
01
Kitchen voicemails, scored
A public test set of spoken restaurant orders and the order lines they should produce: 500 recordings of cooks and managers phoning in or leaving voicemails of orders from sample order guides, with kitchen noise, in at least English and Spanish, each labelled with product, pack size and quantity against a published sample catalog. Score pipelines of speech-to-text plus an AI model that writes the order, line by line.
The briefCredit · unclaimed02
Menu to order guide, scored
A public test of AI tools that draft a restaurant's order guide from its menu: 30 public menus from different cuisines, each paired with order guides built by at least two experienced distributor sales representatives from one published sample catalog, with products, pack sizes and weekly quantities. Score each tool's draft on the products it finds, the products it adds that representatives rejected, and its quantity error.
The briefCredit · unclaimed
Open 01 · the brief
Kitchen voicemails, scored
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set of spoken restaurant orders and the order lines they should produce: 500 recordings of cooks and managers phoning in or leaving voicemails of orders from sample order guides, with kitchen noise, in at least English and Spanish, each labelled with product, pack size and quantity against a published sample catalog. Score pipelines of speech-to-text plus an AI model that writes the order, line by line.
Why it is still open
Choco says it built ground-truth evaluation sets for its own order agents, but they are not public, and it has published no error rates. The nearest public set, ai-coustics' drive-thru intercom recordings posted in September 2026, holds 34 clips from 4 speakers ordering meals, with transcripts but no order lines. Searches of Hugging Face and GitHub found no set of spoken wholesale food orders (November 2026).
Built with what exists
Phone recordings made with volunteers' written consent; speech-to-text models such as ElevenLabs Scribe v3 (see the Note "Fewer word errors") and OpenAI's open Whisper; an AI model such as Claude to turn transcripts into catalog lines; Label Studio, an open labelling tool.
First step
Record 50 orders read by 10 volunteer cooks from a sample order guide, half during service in a working kitchen and half in a quiet room.
Done looks like
500 recordings released under an open licence with their labels, and line-level accuracy (product, pack size and quantity all right) for at least three pipelines, reported separately by noise level and language.
Open 02 · the brief
Menu to order guide, scored
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of AI tools that draft a restaurant's order guide from its menu: 30 public menus from different cuisines, each paired with order guides built by at least two experienced distributor sales representatives from one published sample catalog, with products, pack sizes and weekly quantities. Score each tool's draft on the products it finds, the products it adds that representatives rejected, and its quantity error.
Why it is still open
Vendors already sell this job and report time saved, not accuracy: US Foods reports proposals cut from 3 to 4 hours to 20 minutes, GrubMarket says the work drops from hours to minutes, and Pepper says its assistant builds a mock order guide from a scanned menu in seconds. None publishes how close its guides come to an experienced representative's.
Built with what exists
Public menus from restaurant websites; a sample catalog of foodservice products with pack sizes; AI models that read images and PDFs, such as Claude, Gemini and GPT; a spreadsheet for the representatives' guides; Python for scoring.
First step
Ask two current or retired sales representatives to build order guides for the same five menus from one catalog, and record where they disagree.
Done looks like
30 menu and guide pairs released, the two representatives' agreement with each other reported, and for at least three tools or models the share of representatives' products found, the share of suggested products rejected and the median quantity error.